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 //===----------------------------------------------------------------------===//
13 
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"
25 
26 using namespace clang;
27 using namespace CodeGen;
28 using namespace llvm::omp;
29 
30 namespace {
31 /// Pre(post)-action for different OpenMP constructs specialized for NVPTX.
32 class NVPTXActionTy final : public PrePostActionTy {
33   llvm::FunctionCallee EnterCallee = nullptr;
34   ArrayRef<llvm::Value *> EnterArgs;
35   llvm::FunctionCallee ExitCallee = nullptr;
36   ArrayRef<llvm::Value *> ExitArgs;
37   bool Conditional = false;
38   llvm::BasicBlock *ContBlock = nullptr;
39 
40 public:
41   NVPTXActionTy(llvm::FunctionCallee EnterCallee,
42                 ArrayRef<llvm::Value *> EnterArgs,
43                 llvm::FunctionCallee ExitCallee,
44                 ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false)
45       : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee),
46         ExitArgs(ExitArgs), Conditional(Conditional) {}
47   void Enter(CodeGenFunction &CGF) override {
48     llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs);
49     if (Conditional) {
50       llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes);
51       auto *ThenBlock = CGF.createBasicBlock("omp_if.then");
52       ContBlock = CGF.createBasicBlock("omp_if.end");
53       // Generate the branch (If-stmt)
54       CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock);
55       CGF.EmitBlock(ThenBlock);
56     }
57   }
58   void Done(CodeGenFunction &CGF) {
59     // Emit the rest of blocks/branches
60     CGF.EmitBranch(ContBlock);
61     CGF.EmitBlock(ContBlock, true);
62   }
63   void Exit(CodeGenFunction &CGF) override {
64     CGF.EmitRuntimeCall(ExitCallee, ExitArgs);
65   }
66 };
67 
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.
72 class ExecutionRuntimeModesRAII {
73 private:
74   CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode =
75       CGOpenMPRuntimeGPU::EM_Unknown;
76   CGOpenMPRuntimeGPU::ExecutionMode &ExecMode;
77 
78 public:
79   ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode,
80                             CGOpenMPRuntimeGPU::ExecutionMode EntryMode)
81       : ExecMode(ExecMode) {
82     SavedExecMode = ExecMode;
83     ExecMode = EntryMode;
84   }
85   ~ExecutionRuntimeModesRAII() { ExecMode = SavedExecMode; }
86 };
87 
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.
92 enum MachineConfiguration : unsigned {
93   /// See "llvm/Frontend/OpenMP/OMPGridValues.h" for various related target
94   /// specific Grid Values like GV_Warp_Size, GV_Slot_Size
95 
96   /// Global memory alignment for performance.
97   GlobalMemoryAlignment = 128,
98 };
99 
100 static const ValueDecl *getPrivateItem(const Expr *RefExpr) {
101   RefExpr = RefExpr->IgnoreParens();
102   if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) {
103     const Expr *Base = ASE->getBase()->IgnoreParenImpCasts();
104     while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
105       Base = TempASE->getBase()->IgnoreParenImpCasts();
106     RefExpr = Base;
107   } else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) {
108     const Expr *Base = OASE->getBase()->IgnoreParenImpCasts();
109     while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base))
110       Base = TempOASE->getBase()->IgnoreParenImpCasts();
111     while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
112       Base = TempASE->getBase()->IgnoreParenImpCasts();
113     RefExpr = Base;
114   }
115   RefExpr = RefExpr->IgnoreParenImpCasts();
116   if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr))
117     return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl());
118   const auto *ME = cast<MemberExpr>(RefExpr);
119   return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl());
120 }
121 
122 
123 static RecordDecl *buildRecordForGlobalizedVars(
124     ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls,
125     ArrayRef<const ValueDecl *> EscapedDeclsForTeams,
126     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
127         &MappedDeclsFields, int BufSize) {
128   using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>;
129   if (EscapedDecls.empty() && EscapedDeclsForTeams.empty())
130     return nullptr;
131   SmallVector<VarsDataTy, 4> GlobalizedVars;
132   for (const ValueDecl *D : EscapedDecls)
133     GlobalizedVars.emplace_back(
134         CharUnits::fromQuantity(std::max(
135             C.getDeclAlign(D).getQuantity(),
136             static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))),
137         D);
138   for (const ValueDecl *D : EscapedDeclsForTeams)
139     GlobalizedVars.emplace_back(C.getDeclAlign(D), D);
140   llvm::stable_sort(GlobalizedVars, [](VarsDataTy L, VarsDataTy R) {
141     return L.first > R.first;
142   });
143 
144   // Build struct _globalized_locals_ty {
145   //         /*  globalized vars  */[WarSize] align (max(decl_align,
146   //         GlobalMemoryAlignment))
147   //         /*  globalized vars  */ for EscapedDeclsForTeams
148   //       };
149   RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty");
150   GlobalizedRD->startDefinition();
151   llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped(
152       EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end());
153   for (const auto &Pair : GlobalizedVars) {
154     const ValueDecl *VD = Pair.second;
155     QualType Type = VD->getType();
156     if (Type->isLValueReferenceType())
157       Type = C.getPointerType(Type.getNonReferenceType());
158     else
159       Type = Type.getNonReferenceType();
160     SourceLocation Loc = VD->getLocation();
161     FieldDecl *Field;
162     if (SingleEscaped.count(VD)) {
163       Field = FieldDecl::Create(
164           C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
165           C.getTrivialTypeSourceInfo(Type, SourceLocation()),
166           /*BW=*/nullptr, /*Mutable=*/false,
167           /*InitStyle=*/ICIS_NoInit);
168       Field->setAccess(AS_public);
169       if (VD->hasAttrs()) {
170         for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
171              E(VD->getAttrs().end());
172              I != E; ++I)
173           Field->addAttr(*I);
174       }
175     } else {
176       llvm::APInt ArraySize(32, BufSize);
177       Type = C.getConstantArrayType(Type, ArraySize, nullptr, ArrayType::Normal,
178                                     0);
179       Field = FieldDecl::Create(
180           C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
181           C.getTrivialTypeSourceInfo(Type, SourceLocation()),
182           /*BW=*/nullptr, /*Mutable=*/false,
183           /*InitStyle=*/ICIS_NoInit);
184       Field->setAccess(AS_public);
185       llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(),
186                                      static_cast<CharUnits::QuantityType>(
187                                          GlobalMemoryAlignment)));
188       Field->addAttr(AlignedAttr::CreateImplicit(
189           C, /*IsAlignmentExpr=*/true,
190           IntegerLiteral::Create(C, Align,
191                                  C.getIntTypeForBitwidth(32, /*Signed=*/0),
192                                  SourceLocation()),
193           {}, AlignedAttr::GNU_aligned));
194     }
195     GlobalizedRD->addDecl(Field);
196     MappedDeclsFields.try_emplace(VD, Field);
197   }
198   GlobalizedRD->completeDefinition();
199   return GlobalizedRD;
200 }
201 
202 /// Get the list of variables that can escape their declaration context.
203 class CheckVarsEscapingDeclContext final
204     : public ConstStmtVisitor<CheckVarsEscapingDeclContext> {
205   CodeGenFunction &CGF;
206   llvm::SetVector<const ValueDecl *> EscapedDecls;
207   llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls;
208   llvm::SetVector<const ValueDecl *> DelayedVariableLengthDecls;
209   llvm::SmallPtrSet<const Decl *, 4> EscapedParameters;
210   RecordDecl *GlobalizedRD = nullptr;
211   llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
212   bool AllEscaped = false;
213   bool IsForCombinedParallelRegion = false;
214 
215   void markAsEscaped(const ValueDecl *VD) {
216     // Do not globalize declare target variables.
217     if (!isa<VarDecl>(VD) ||
218         OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD))
219       return;
220     VD = cast<ValueDecl>(VD->getCanonicalDecl());
221     // Use user-specified allocation.
222     if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>())
223       return;
224     // Variables captured by value must be globalized.
225     bool IsCaptured = false;
226     if (auto *CSI = CGF.CapturedStmtInfo) {
227       if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) {
228         // Check if need to capture the variable that was already captured by
229         // value in the outer region.
230         IsCaptured = true;
231         if (!IsForCombinedParallelRegion) {
232           if (!FD->hasAttrs())
233             return;
234           const auto *Attr = FD->getAttr<OMPCaptureKindAttr>();
235           if (!Attr)
236             return;
237           if (((Attr->getCaptureKind() != OMPC_map) &&
238                !isOpenMPPrivate(Attr->getCaptureKind())) ||
239               ((Attr->getCaptureKind() == OMPC_map) &&
240                !FD->getType()->isAnyPointerType()))
241             return;
242         }
243         if (!FD->getType()->isReferenceType()) {
244           assert(!VD->getType()->isVariablyModifiedType() &&
245                  "Parameter captured by value with variably modified type");
246           EscapedParameters.insert(VD);
247         } else if (!IsForCombinedParallelRegion) {
248           return;
249         }
250       }
251     }
252     if ((!CGF.CapturedStmtInfo ||
253          (IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) &&
254         VD->getType()->isReferenceType())
255       // Do not globalize variables with reference type.
256       return;
257     if (VD->getType()->isVariablyModifiedType()) {
258       // If not captured at the target region level then mark the escaped
259       // variable as delayed.
260       if (IsCaptured)
261         EscapedVariableLengthDecls.insert(VD);
262       else
263         DelayedVariableLengthDecls.insert(VD);
264     } else
265       EscapedDecls.insert(VD);
266   }
267 
268   void VisitValueDecl(const ValueDecl *VD) {
269     if (VD->getType()->isLValueReferenceType())
270       markAsEscaped(VD);
271     if (const auto *VarD = dyn_cast<VarDecl>(VD)) {
272       if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) {
273         const bool SavedAllEscaped = AllEscaped;
274         AllEscaped = VD->getType()->isLValueReferenceType();
275         Visit(VarD->getInit());
276         AllEscaped = SavedAllEscaped;
277       }
278     }
279   }
280   void VisitOpenMPCapturedStmt(const CapturedStmt *S,
281                                ArrayRef<OMPClause *> Clauses,
282                                bool IsCombinedParallelRegion) {
283     if (!S)
284       return;
285     for (const CapturedStmt::Capture &C : S->captures()) {
286       if (C.capturesVariable() && !C.capturesVariableByCopy()) {
287         const ValueDecl *VD = C.getCapturedVar();
288         bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion;
289         if (IsCombinedParallelRegion) {
290           // Check if the variable is privatized in the combined construct and
291           // those private copies must be shared in the inner parallel
292           // directive.
293           IsForCombinedParallelRegion = false;
294           for (const OMPClause *C : Clauses) {
295             if (!isOpenMPPrivate(C->getClauseKind()) ||
296                 C->getClauseKind() == OMPC_reduction ||
297                 C->getClauseKind() == OMPC_linear ||
298                 C->getClauseKind() == OMPC_private)
299               continue;
300             ArrayRef<const Expr *> Vars;
301             if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C))
302               Vars = PC->getVarRefs();
303             else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C))
304               Vars = PC->getVarRefs();
305             else
306               llvm_unreachable("Unexpected clause.");
307             for (const auto *E : Vars) {
308               const Decl *D =
309                   cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl();
310               if (D == VD->getCanonicalDecl()) {
311                 IsForCombinedParallelRegion = true;
312                 break;
313               }
314             }
315             if (IsForCombinedParallelRegion)
316               break;
317           }
318         }
319         markAsEscaped(VD);
320         if (isa<OMPCapturedExprDecl>(VD))
321           VisitValueDecl(VD);
322         IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion;
323       }
324     }
325   }
326 
327   void buildRecordForGlobalizedVars(bool IsInTTDRegion) {
328     assert(!GlobalizedRD &&
329            "Record for globalized variables is built already.");
330     ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams;
331     unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size;
332     if (IsInTTDRegion)
333       EscapedDeclsForTeams = EscapedDecls.getArrayRef();
334     else
335       EscapedDeclsForParallel = EscapedDecls.getArrayRef();
336     GlobalizedRD = ::buildRecordForGlobalizedVars(
337         CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams,
338         MappedDeclsFields, WarpSize);
339   }
340 
341 public:
342   CheckVarsEscapingDeclContext(CodeGenFunction &CGF,
343                                ArrayRef<const ValueDecl *> TeamsReductions)
344       : CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) {
345   }
346   virtual ~CheckVarsEscapingDeclContext() = default;
347   void VisitDeclStmt(const DeclStmt *S) {
348     if (!S)
349       return;
350     for (const Decl *D : S->decls())
351       if (const auto *VD = dyn_cast_or_null<ValueDecl>(D))
352         VisitValueDecl(VD);
353   }
354   void VisitOMPExecutableDirective(const OMPExecutableDirective *D) {
355     if (!D)
356       return;
357     if (!D->hasAssociatedStmt())
358       return;
359     if (const auto *S =
360             dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) {
361       // Do not analyze directives that do not actually require capturing,
362       // like `omp for` or `omp simd` directives.
363       llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
364       getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind());
365       if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) {
366         VisitStmt(S->getCapturedStmt());
367         return;
368       }
369       VisitOpenMPCapturedStmt(
370           S, D->clauses(),
371           CaptureRegions.back() == OMPD_parallel &&
372               isOpenMPDistributeDirective(D->getDirectiveKind()));
373     }
374   }
375   void VisitCapturedStmt(const CapturedStmt *S) {
376     if (!S)
377       return;
378     for (const CapturedStmt::Capture &C : S->captures()) {
379       if (C.capturesVariable() && !C.capturesVariableByCopy()) {
380         const ValueDecl *VD = C.getCapturedVar();
381         markAsEscaped(VD);
382         if (isa<OMPCapturedExprDecl>(VD))
383           VisitValueDecl(VD);
384       }
385     }
386   }
387   void VisitLambdaExpr(const LambdaExpr *E) {
388     if (!E)
389       return;
390     for (const LambdaCapture &C : E->captures()) {
391       if (C.capturesVariable()) {
392         if (C.getCaptureKind() == LCK_ByRef) {
393           const ValueDecl *VD = C.getCapturedVar();
394           markAsEscaped(VD);
395           if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD))
396             VisitValueDecl(VD);
397         }
398       }
399     }
400   }
401   void VisitBlockExpr(const BlockExpr *E) {
402     if (!E)
403       return;
404     for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) {
405       if (C.isByRef()) {
406         const VarDecl *VD = C.getVariable();
407         markAsEscaped(VD);
408         if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture())
409           VisitValueDecl(VD);
410       }
411     }
412   }
413   void VisitCallExpr(const CallExpr *E) {
414     if (!E)
415       return;
416     for (const Expr *Arg : E->arguments()) {
417       if (!Arg)
418         continue;
419       if (Arg->isLValue()) {
420         const bool SavedAllEscaped = AllEscaped;
421         AllEscaped = true;
422         Visit(Arg);
423         AllEscaped = SavedAllEscaped;
424       } else {
425         Visit(Arg);
426       }
427     }
428     Visit(E->getCallee());
429   }
430   void VisitDeclRefExpr(const DeclRefExpr *E) {
431     if (!E)
432       return;
433     const ValueDecl *VD = E->getDecl();
434     if (AllEscaped)
435       markAsEscaped(VD);
436     if (isa<OMPCapturedExprDecl>(VD))
437       VisitValueDecl(VD);
438     else if (VD->isInitCapture())
439       VisitValueDecl(VD);
440   }
441   void VisitUnaryOperator(const UnaryOperator *E) {
442     if (!E)
443       return;
444     if (E->getOpcode() == UO_AddrOf) {
445       const bool SavedAllEscaped = AllEscaped;
446       AllEscaped = true;
447       Visit(E->getSubExpr());
448       AllEscaped = SavedAllEscaped;
449     } else {
450       Visit(E->getSubExpr());
451     }
452   }
453   void VisitImplicitCastExpr(const ImplicitCastExpr *E) {
454     if (!E)
455       return;
456     if (E->getCastKind() == CK_ArrayToPointerDecay) {
457       const bool SavedAllEscaped = AllEscaped;
458       AllEscaped = true;
459       Visit(E->getSubExpr());
460       AllEscaped = SavedAllEscaped;
461     } else {
462       Visit(E->getSubExpr());
463     }
464   }
465   void VisitExpr(const Expr *E) {
466     if (!E)
467       return;
468     bool SavedAllEscaped = AllEscaped;
469     if (!E->isLValue())
470       AllEscaped = false;
471     for (const Stmt *Child : E->children())
472       if (Child)
473         Visit(Child);
474     AllEscaped = SavedAllEscaped;
475   }
476   void VisitStmt(const Stmt *S) {
477     if (!S)
478       return;
479     for (const Stmt *Child : S->children())
480       if (Child)
481         Visit(Child);
482   }
483 
484   /// Returns the record that handles all the escaped local variables and used
485   /// instead of their original storage.
486   const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) {
487     if (!GlobalizedRD)
488       buildRecordForGlobalizedVars(IsInTTDRegion);
489     return GlobalizedRD;
490   }
491 
492   /// Returns the field in the globalized record for the escaped variable.
493   const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const {
494     assert(GlobalizedRD &&
495            "Record for globalized variables must be generated already.");
496     return MappedDeclsFields.lookup(VD);
497   }
498 
499   /// Returns the list of the escaped local variables/parameters.
500   ArrayRef<const ValueDecl *> getEscapedDecls() const {
501     return EscapedDecls.getArrayRef();
502   }
503 
504   /// Checks if the escaped local variable is actually a parameter passed by
505   /// value.
506   const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const {
507     return EscapedParameters;
508   }
509 
510   /// Returns the list of the escaped variables with the variably modified
511   /// types.
512   ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const {
513     return EscapedVariableLengthDecls.getArrayRef();
514   }
515 
516   /// Returns the list of the delayed variables with the variably modified
517   /// types.
518   ArrayRef<const ValueDecl *> getDelayedVariableLengthDecls() const {
519     return DelayedVariableLengthDecls.getArrayRef();
520   }
521 };
522 } // anonymous namespace
523 
524 /// Get the id of the warp in the block.
525 /// We assume that the warp size is 32, which is always the case
526 /// on the NVPTX device, to generate more efficient code.
527 static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) {
528   CGBuilderTy &Bld = CGF.Builder;
529   unsigned LaneIDBits =
530       llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size);
531   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
532   return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id");
533 }
534 
535 /// Get the id of the current lane in the Warp.
536 /// We assume that the warp size is 32, which is always the case
537 /// on the NVPTX device, to generate more efficient code.
538 static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) {
539   CGBuilderTy &Bld = CGF.Builder;
540   unsigned LaneIDBits =
541       llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size);
542   assert(LaneIDBits < 32 && "Invalid LaneIDBits size in NVPTX device.");
543   unsigned LaneIDMask = ~0u >> (32u - LaneIDBits);
544   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
545   return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask),
546                        "nvptx_lane_id");
547 }
548 
549 CGOpenMPRuntimeGPU::ExecutionMode
550 CGOpenMPRuntimeGPU::getExecutionMode() const {
551   return CurrentExecutionMode;
552 }
553 
554 static CGOpenMPRuntimeGPU::DataSharingMode
555 getDataSharingMode(CodeGenModule &CGM) {
556   return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeGPU::CUDA
557                                           : CGOpenMPRuntimeGPU::Generic;
558 }
559 
560 /// Check for inner (nested) SPMD construct, if any
561 static bool hasNestedSPMDDirective(ASTContext &Ctx,
562                                    const OMPExecutableDirective &D) {
563   const auto *CS = D.getInnermostCapturedStmt();
564   const auto *Body =
565       CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
566   const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
567 
568   if (const auto *NestedDir =
569           dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
570     OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
571     switch (D.getDirectiveKind()) {
572     case OMPD_target:
573       if (isOpenMPParallelDirective(DKind))
574         return true;
575       if (DKind == OMPD_teams) {
576         Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
577             /*IgnoreCaptured=*/true);
578         if (!Body)
579           return false;
580         ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
581         if (const auto *NND =
582                 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
583           DKind = NND->getDirectiveKind();
584           if (isOpenMPParallelDirective(DKind))
585             return true;
586         }
587       }
588       return false;
589     case OMPD_target_teams:
590       return isOpenMPParallelDirective(DKind);
591     case OMPD_target_simd:
592     case OMPD_target_parallel:
593     case OMPD_target_parallel_for:
594     case OMPD_target_parallel_for_simd:
595     case OMPD_target_teams_distribute:
596     case OMPD_target_teams_distribute_simd:
597     case OMPD_target_teams_distribute_parallel_for:
598     case OMPD_target_teams_distribute_parallel_for_simd:
599     case OMPD_parallel:
600     case OMPD_for:
601     case OMPD_parallel_for:
602     case OMPD_parallel_master:
603     case OMPD_parallel_sections:
604     case OMPD_for_simd:
605     case OMPD_parallel_for_simd:
606     case OMPD_cancel:
607     case OMPD_cancellation_point:
608     case OMPD_ordered:
609     case OMPD_threadprivate:
610     case OMPD_allocate:
611     case OMPD_task:
612     case OMPD_simd:
613     case OMPD_sections:
614     case OMPD_section:
615     case OMPD_single:
616     case OMPD_master:
617     case OMPD_critical:
618     case OMPD_taskyield:
619     case OMPD_barrier:
620     case OMPD_taskwait:
621     case OMPD_taskgroup:
622     case OMPD_atomic:
623     case OMPD_flush:
624     case OMPD_depobj:
625     case OMPD_scan:
626     case OMPD_teams:
627     case OMPD_target_data:
628     case OMPD_target_exit_data:
629     case OMPD_target_enter_data:
630     case OMPD_distribute:
631     case OMPD_distribute_simd:
632     case OMPD_distribute_parallel_for:
633     case OMPD_distribute_parallel_for_simd:
634     case OMPD_teams_distribute:
635     case OMPD_teams_distribute_simd:
636     case OMPD_teams_distribute_parallel_for:
637     case OMPD_teams_distribute_parallel_for_simd:
638     case OMPD_target_update:
639     case OMPD_declare_simd:
640     case OMPD_declare_variant:
641     case OMPD_begin_declare_variant:
642     case OMPD_end_declare_variant:
643     case OMPD_declare_target:
644     case OMPD_end_declare_target:
645     case OMPD_declare_reduction:
646     case OMPD_declare_mapper:
647     case OMPD_taskloop:
648     case OMPD_taskloop_simd:
649     case OMPD_master_taskloop:
650     case OMPD_master_taskloop_simd:
651     case OMPD_parallel_master_taskloop:
652     case OMPD_parallel_master_taskloop_simd:
653     case OMPD_requires:
654     case OMPD_unknown:
655     default:
656       llvm_unreachable("Unexpected directive.");
657     }
658   }
659 
660   return false;
661 }
662 
663 static bool supportsSPMDExecutionMode(ASTContext &Ctx,
664                                       const OMPExecutableDirective &D) {
665   OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
666   switch (DirectiveKind) {
667   case OMPD_target:
668   case OMPD_target_teams:
669     return hasNestedSPMDDirective(Ctx, D);
670   case OMPD_target_teams_loop:
671   case OMPD_target_parallel_loop:
672   case OMPD_target_parallel:
673   case OMPD_target_parallel_for:
674   case OMPD_target_parallel_for_simd:
675   case OMPD_target_teams_distribute_parallel_for:
676   case OMPD_target_teams_distribute_parallel_for_simd:
677   case OMPD_target_simd:
678   case OMPD_target_teams_distribute_simd:
679     return true;
680   case OMPD_target_teams_distribute:
681     return false;
682   case OMPD_parallel:
683   case OMPD_for:
684   case OMPD_parallel_for:
685   case OMPD_parallel_master:
686   case OMPD_parallel_sections:
687   case OMPD_for_simd:
688   case OMPD_parallel_for_simd:
689   case OMPD_cancel:
690   case OMPD_cancellation_point:
691   case OMPD_ordered:
692   case OMPD_threadprivate:
693   case OMPD_allocate:
694   case OMPD_task:
695   case OMPD_simd:
696   case OMPD_sections:
697   case OMPD_section:
698   case OMPD_single:
699   case OMPD_master:
700   case OMPD_critical:
701   case OMPD_taskyield:
702   case OMPD_barrier:
703   case OMPD_taskwait:
704   case OMPD_taskgroup:
705   case OMPD_atomic:
706   case OMPD_flush:
707   case OMPD_depobj:
708   case OMPD_scan:
709   case OMPD_teams:
710   case OMPD_target_data:
711   case OMPD_target_exit_data:
712   case OMPD_target_enter_data:
713   case OMPD_distribute:
714   case OMPD_distribute_simd:
715   case OMPD_distribute_parallel_for:
716   case OMPD_distribute_parallel_for_simd:
717   case OMPD_teams_distribute:
718   case OMPD_teams_distribute_simd:
719   case OMPD_teams_distribute_parallel_for:
720   case OMPD_teams_distribute_parallel_for_simd:
721   case OMPD_target_update:
722   case OMPD_declare_simd:
723   case OMPD_declare_variant:
724   case OMPD_begin_declare_variant:
725   case OMPD_end_declare_variant:
726   case OMPD_declare_target:
727   case OMPD_end_declare_target:
728   case OMPD_declare_reduction:
729   case OMPD_declare_mapper:
730   case OMPD_taskloop:
731   case OMPD_taskloop_simd:
732   case OMPD_master_taskloop:
733   case OMPD_master_taskloop_simd:
734   case OMPD_parallel_master_taskloop:
735   case OMPD_parallel_master_taskloop_simd:
736   case OMPD_requires:
737   case OMPD_unknown:
738   default:
739     break;
740   }
741   llvm_unreachable(
742       "Unknown programming model for OpenMP directive on NVPTX target.");
743 }
744 
745 void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D,
746                                              StringRef ParentName,
747                                              llvm::Function *&OutlinedFn,
748                                              llvm::Constant *&OutlinedFnID,
749                                              bool IsOffloadEntry,
750                                              const RegionCodeGenTy &CodeGen) {
751   ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_NonSPMD);
752   EntryFunctionState EST;
753   WrapperFunctionsMap.clear();
754 
755   // Emit target region as a standalone region.
756   class NVPTXPrePostActionTy : public PrePostActionTy {
757     CGOpenMPRuntimeGPU::EntryFunctionState &EST;
758 
759   public:
760     NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST)
761         : EST(EST) {}
762     void Enter(CodeGenFunction &CGF) override {
763       auto &RT =
764           static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
765       RT.emitKernelInit(CGF, EST, /* IsSPMD */ false);
766       // Skip target region initialization.
767       RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
768     }
769     void Exit(CodeGenFunction &CGF) override {
770       auto &RT =
771           static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
772       RT.clearLocThreadIdInsertPt(CGF);
773       RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ false);
774     }
775   } Action(EST);
776   CodeGen.setAction(Action);
777   IsInTTDRegion = true;
778   emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
779                                    IsOffloadEntry, CodeGen);
780   IsInTTDRegion = false;
781 }
782 
783 void CGOpenMPRuntimeGPU::emitKernelInit(CodeGenFunction &CGF,
784                                         EntryFunctionState &EST, bool IsSPMD) {
785   CGBuilderTy &Bld = CGF.Builder;
786   Bld.restoreIP(OMPBuilder.createTargetInit(Bld, IsSPMD));
787   if (!IsSPMD)
788     emitGenericVarsProlog(CGF, EST.Loc);
789 }
790 
791 void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF,
792                                           EntryFunctionState &EST,
793                                           bool IsSPMD) {
794   if (!IsSPMD)
795     emitGenericVarsEpilog(CGF);
796 
797   CGBuilderTy &Bld = CGF.Builder;
798   OMPBuilder.createTargetDeinit(Bld, IsSPMD);
799 }
800 
801 void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D,
802                                           StringRef ParentName,
803                                           llvm::Function *&OutlinedFn,
804                                           llvm::Constant *&OutlinedFnID,
805                                           bool IsOffloadEntry,
806                                           const RegionCodeGenTy &CodeGen) {
807   ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_SPMD);
808   EntryFunctionState EST;
809 
810   // Emit target region as a standalone region.
811   class NVPTXPrePostActionTy : public PrePostActionTy {
812     CGOpenMPRuntimeGPU &RT;
813     CGOpenMPRuntimeGPU::EntryFunctionState &EST;
814 
815   public:
816     NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT,
817                          CGOpenMPRuntimeGPU::EntryFunctionState &EST)
818         : RT(RT), EST(EST) {}
819     void Enter(CodeGenFunction &CGF) override {
820       RT.emitKernelInit(CGF, EST, /* IsSPMD */ true);
821       // Skip target region initialization.
822       RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
823     }
824     void Exit(CodeGenFunction &CGF) override {
825       RT.clearLocThreadIdInsertPt(CGF);
826       RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true);
827     }
828   } Action(*this, EST);
829   CodeGen.setAction(Action);
830   IsInTTDRegion = true;
831   emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
832                                    IsOffloadEntry, CodeGen);
833   IsInTTDRegion = false;
834 }
835 
836 // Create a unique global variable to indicate the execution mode of this target
837 // region. The execution mode is either 'generic', or 'spmd' depending on the
838 // target directive. This variable is picked up by the offload library to setup
839 // the device appropriately before kernel launch. If the execution mode is
840 // 'generic', the runtime reserves one warp for the master, otherwise, all
841 // warps participate in parallel work.
842 static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name,
843                                      bool Mode) {
844   auto *GVMode = new llvm::GlobalVariable(
845       CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true,
846       llvm::GlobalValue::WeakAnyLinkage,
847       llvm::ConstantInt::get(CGM.Int8Ty, Mode ? OMP_TGT_EXEC_MODE_SPMD
848                                               : OMP_TGT_EXEC_MODE_GENERIC),
849       Twine(Name, "_exec_mode"));
850   GVMode->setVisibility(llvm::GlobalVariable::ProtectedVisibility);
851   CGM.addCompilerUsedGlobal(GVMode);
852 }
853 
854 void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction(
855     const OMPExecutableDirective &D, StringRef ParentName,
856     llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
857     bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
858   if (!IsOffloadEntry) // Nothing to do.
859     return;
860 
861   assert(!ParentName.empty() && "Invalid target region parent name!");
862 
863   bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D);
864   if (Mode)
865     emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
866                    CodeGen);
867   else
868     emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
869                       CodeGen);
870 
871   setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode);
872 }
873 
874 CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM)
875     : CGOpenMPRuntime(CGM) {
876   llvm::OpenMPIRBuilderConfig Config(CGM.getLangOpts().OpenMPIsTargetDevice,
877                                      isGPU(), hasRequiresUnifiedSharedMemory(),
878                                      CGM.getLangOpts().OpenMPOffloadMandatory);
879   OMPBuilder.setConfig(Config);
880 
881   if (!CGM.getLangOpts().OpenMPIsTargetDevice)
882     llvm_unreachable("OpenMP can only handle device code.");
883 
884   llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder();
885   if (CGM.getLangOpts().NoGPULib || CGM.getLangOpts().OMPHostIRFile.empty())
886     return;
887 
888   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTargetDebug,
889                               "__omp_rtl_debug_kind");
890   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTeamSubscription,
891                               "__omp_rtl_assume_teams_oversubscription");
892   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPThreadSubscription,
893                               "__omp_rtl_assume_threads_oversubscription");
894   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoThreadState,
895                               "__omp_rtl_assume_no_thread_state");
896   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoNestedParallelism,
897                               "__omp_rtl_assume_no_nested_parallelism");
898 }
899 
900 void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF,
901                                               ProcBindKind ProcBind,
902                                               SourceLocation Loc) {
903   // Do nothing in case of SPMD mode and L0 parallel.
904   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
905     return;
906 
907   CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc);
908 }
909 
910 void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF,
911                                                 llvm::Value *NumThreads,
912                                                 SourceLocation Loc) {
913   // Nothing to do.
914 }
915 
916 void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF,
917                                               const Expr *NumTeams,
918                                               const Expr *ThreadLimit,
919                                               SourceLocation Loc) {}
920 
921 llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction(
922     CodeGenFunction &CGF, const OMPExecutableDirective &D,
923     const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind,
924     const RegionCodeGenTy &CodeGen) {
925   // Emit target region as a standalone region.
926   bool PrevIsInTTDRegion = IsInTTDRegion;
927   IsInTTDRegion = false;
928   auto *OutlinedFun =
929       cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction(
930           CGF, D, ThreadIDVar, InnermostKind, CodeGen));
931   IsInTTDRegion = PrevIsInTTDRegion;
932   if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) {
933     llvm::Function *WrapperFun =
934         createParallelDataSharingWrapper(OutlinedFun, D);
935     WrapperFunctionsMap[OutlinedFun] = WrapperFun;
936   }
937 
938   return OutlinedFun;
939 }
940 
941 /// Get list of lastprivate variables from the teams distribute ... or
942 /// teams {distribute ...} directives.
943 static void
944 getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D,
945                              llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
946   assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&
947          "expected teams directive.");
948   const OMPExecutableDirective *Dir = &D;
949   if (!isOpenMPDistributeDirective(D.getDirectiveKind())) {
950     if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild(
951             Ctx,
952             D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(
953                 /*IgnoreCaptured=*/true))) {
954       Dir = dyn_cast_or_null<OMPExecutableDirective>(S);
955       if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind()))
956         Dir = nullptr;
957     }
958   }
959   if (!Dir)
960     return;
961   for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) {
962     for (const Expr *E : C->getVarRefs())
963       Vars.push_back(getPrivateItem(E));
964   }
965 }
966 
967 /// Get list of reduction variables from the teams ... directives.
968 static void
969 getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D,
970                       llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
971   assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&
972          "expected teams directive.");
973   for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
974     for (const Expr *E : C->privates())
975       Vars.push_back(getPrivateItem(E));
976   }
977 }
978 
979 llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction(
980     CodeGenFunction &CGF, const OMPExecutableDirective &D,
981     const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind,
982     const RegionCodeGenTy &CodeGen) {
983   SourceLocation Loc = D.getBeginLoc();
984 
985   const RecordDecl *GlobalizedRD = nullptr;
986   llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions;
987   llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
988   unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size;
989   // Globalize team reductions variable unconditionally in all modes.
990   if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
991     getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions);
992   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
993     getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions);
994     if (!LastPrivatesReductions.empty()) {
995       GlobalizedRD = ::buildRecordForGlobalizedVars(
996           CGM.getContext(), std::nullopt, LastPrivatesReductions,
997           MappedDeclsFields, WarpSize);
998     }
999   } else if (!LastPrivatesReductions.empty()) {
1000     assert(!TeamAndReductions.first &&
1001            "Previous team declaration is not expected.");
1002     TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl();
1003     std::swap(TeamAndReductions.second, LastPrivatesReductions);
1004   }
1005 
1006   // Emit target region as a standalone region.
1007   class NVPTXPrePostActionTy : public PrePostActionTy {
1008     SourceLocation &Loc;
1009     const RecordDecl *GlobalizedRD;
1010     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
1011         &MappedDeclsFields;
1012 
1013   public:
1014     NVPTXPrePostActionTy(
1015         SourceLocation &Loc, const RecordDecl *GlobalizedRD,
1016         llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
1017             &MappedDeclsFields)
1018         : Loc(Loc), GlobalizedRD(GlobalizedRD),
1019           MappedDeclsFields(MappedDeclsFields) {}
1020     void Enter(CodeGenFunction &CGF) override {
1021       auto &Rt =
1022           static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1023       if (GlobalizedRD) {
1024         auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
1025         I->getSecond().MappedParams =
1026             std::make_unique<CodeGenFunction::OMPMapVars>();
1027         DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
1028         for (const auto &Pair : MappedDeclsFields) {
1029           assert(Pair.getFirst()->isCanonicalDecl() &&
1030                  "Expected canonical declaration");
1031           Data.insert(std::make_pair(Pair.getFirst(), MappedVarData()));
1032         }
1033       }
1034       Rt.emitGenericVarsProlog(CGF, Loc);
1035     }
1036     void Exit(CodeGenFunction &CGF) override {
1037       static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
1038           .emitGenericVarsEpilog(CGF);
1039     }
1040   } Action(Loc, GlobalizedRD, MappedDeclsFields);
1041   CodeGen.setAction(Action);
1042   llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction(
1043       CGF, D, ThreadIDVar, InnermostKind, CodeGen);
1044 
1045   return OutlinedFun;
1046 }
1047 
1048 void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF,
1049                                                  SourceLocation Loc,
1050                                                  bool WithSPMDCheck) {
1051   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic &&
1052       getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
1053     return;
1054 
1055   CGBuilderTy &Bld = CGF.Builder;
1056 
1057   const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
1058   if (I == FunctionGlobalizedDecls.end())
1059     return;
1060 
1061   for (auto &Rec : I->getSecond().LocalVarData) {
1062     const auto *VD = cast<VarDecl>(Rec.first);
1063     bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first);
1064     QualType VarTy = VD->getType();
1065 
1066     // Get the local allocation of a firstprivate variable before sharing
1067     llvm::Value *ParValue;
1068     if (EscapedParam) {
1069       LValue ParLVal =
1070           CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
1071       ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc);
1072     }
1073 
1074     // Allocate space for the variable to be globalized
1075     llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())};
1076     llvm::CallBase *VoidPtr =
1077         CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1078                                 CGM.getModule(), OMPRTL___kmpc_alloc_shared),
1079                             AllocArgs, VD->getName());
1080     // FIXME: We should use the variables actual alignment as an argument.
1081     VoidPtr->addRetAttr(llvm::Attribute::get(
1082         CGM.getLLVMContext(), llvm::Attribute::Alignment,
1083         CGM.getContext().getTargetInfo().getNewAlign() / 8));
1084 
1085     // Cast the void pointer and get the address of the globalized variable.
1086     llvm::PointerType *VarPtrTy = CGF.ConvertTypeForMem(VarTy)->getPointerTo();
1087     llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
1088         VoidPtr, VarPtrTy, VD->getName() + "_on_stack");
1089     LValue VarAddr = CGF.MakeNaturalAlignAddrLValue(CastedVoidPtr, VarTy);
1090     Rec.second.PrivateAddr = VarAddr.getAddress(CGF);
1091     Rec.second.GlobalizedVal = VoidPtr;
1092 
1093     // Assign the local allocation to the newly globalized location.
1094     if (EscapedParam) {
1095       CGF.EmitStoreOfScalar(ParValue, VarAddr);
1096       I->getSecond().MappedParams->setVarAddr(CGF, VD, VarAddr.getAddress(CGF));
1097     }
1098     if (auto *DI = CGF.getDebugInfo())
1099       VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(VD->getLocation()));
1100   }
1101 
1102   for (const auto *ValueD : I->getSecond().EscapedVariableLengthDecls) {
1103     const auto *VD = cast<VarDecl>(ValueD);
1104     std::pair<llvm::Value *, llvm::Value *> AddrSizePair =
1105         getKmpcAllocShared(CGF, VD);
1106     I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(AddrSizePair);
1107     LValue Base = CGF.MakeAddrLValue(AddrSizePair.first, VD->getType(),
1108                                      CGM.getContext().getDeclAlign(VD),
1109                                      AlignmentSource::Decl);
1110     I->getSecond().MappedParams->setVarAddr(CGF, VD, Base.getAddress(CGF));
1111   }
1112   I->getSecond().MappedParams->apply(CGF);
1113 }
1114 
1115 bool CGOpenMPRuntimeGPU::isDelayedVariableLengthDecl(CodeGenFunction &CGF,
1116                                                      const VarDecl *VD) const {
1117   const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
1118   if (I == FunctionGlobalizedDecls.end())
1119     return false;
1120 
1121   // Check variable declaration is delayed:
1122   return llvm::is_contained(I->getSecond().DelayedVariableLengthDecls, VD);
1123 }
1124 
1125 std::pair<llvm::Value *, llvm::Value *>
1126 CGOpenMPRuntimeGPU::getKmpcAllocShared(CodeGenFunction &CGF,
1127                                        const VarDecl *VD) {
1128   CGBuilderTy &Bld = CGF.Builder;
1129 
1130   // Compute size and alignment.
1131   llvm::Value *Size = CGF.getTypeSize(VD->getType());
1132   CharUnits Align = CGM.getContext().getDeclAlign(VD);
1133   Size = Bld.CreateNUWAdd(
1134       Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1));
1135   llvm::Value *AlignVal =
1136       llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity());
1137   Size = Bld.CreateUDiv(Size, AlignVal);
1138   Size = Bld.CreateNUWMul(Size, AlignVal);
1139 
1140   // Allocate space for this VLA object to be globalized.
1141   llvm::Value *AllocArgs[] = {Size};
1142   llvm::CallBase *VoidPtr =
1143       CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1144                               CGM.getModule(), OMPRTL___kmpc_alloc_shared),
1145                           AllocArgs, VD->getName());
1146   VoidPtr->addRetAttr(llvm::Attribute::get(
1147       CGM.getLLVMContext(), llvm::Attribute::Alignment, Align.getQuantity()));
1148 
1149   return std::make_pair(VoidPtr, Size);
1150 }
1151 
1152 void CGOpenMPRuntimeGPU::getKmpcFreeShared(
1153     CodeGenFunction &CGF,
1154     const std::pair<llvm::Value *, llvm::Value *> &AddrSizePair) {
1155   // Deallocate the memory for each globalized VLA object
1156   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1157                           CGM.getModule(), OMPRTL___kmpc_free_shared),
1158                       {AddrSizePair.first, AddrSizePair.second});
1159 }
1160 
1161 void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF,
1162                                                  bool WithSPMDCheck) {
1163   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic &&
1164       getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
1165     return;
1166 
1167   const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
1168   if (I != FunctionGlobalizedDecls.end()) {
1169     // Deallocate the memory for each globalized VLA object that was
1170     // globalized in the prolog (i.e. emitGenericVarsProlog).
1171     for (const auto &AddrSizePair :
1172          llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) {
1173       CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1174                               CGM.getModule(), OMPRTL___kmpc_free_shared),
1175                           {AddrSizePair.first, AddrSizePair.second});
1176     }
1177     // Deallocate the memory for each globalized value
1178     for (auto &Rec : llvm::reverse(I->getSecond().LocalVarData)) {
1179       const auto *VD = cast<VarDecl>(Rec.first);
1180       I->getSecond().MappedParams->restore(CGF);
1181 
1182       llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal,
1183                                  CGF.getTypeSize(VD->getType())};
1184       CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1185                               CGM.getModule(), OMPRTL___kmpc_free_shared),
1186                           FreeArgs);
1187     }
1188   }
1189 }
1190 
1191 void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF,
1192                                          const OMPExecutableDirective &D,
1193                                          SourceLocation Loc,
1194                                          llvm::Function *OutlinedFn,
1195                                          ArrayRef<llvm::Value *> CapturedVars) {
1196   if (!CGF.HaveInsertPoint())
1197     return;
1198 
1199   Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
1200                                                       /*Name=*/".zero.addr");
1201   CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr);
1202   llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
1203   OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer());
1204   OutlinedFnArgs.push_back(ZeroAddr.getPointer());
1205   OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
1206   emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs);
1207 }
1208 
1209 void CGOpenMPRuntimeGPU::emitParallelCall(CodeGenFunction &CGF,
1210                                           SourceLocation Loc,
1211                                           llvm::Function *OutlinedFn,
1212                                           ArrayRef<llvm::Value *> CapturedVars,
1213                                           const Expr *IfCond,
1214                                           llvm::Value *NumThreads) {
1215   if (!CGF.HaveInsertPoint())
1216     return;
1217 
1218   auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond,
1219                         NumThreads](CodeGenFunction &CGF,
1220                                     PrePostActionTy &Action) {
1221     CGBuilderTy &Bld = CGF.Builder;
1222     llvm::Value *NumThreadsVal = NumThreads;
1223     llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn];
1224     llvm::Value *ID = llvm::ConstantPointerNull::get(CGM.Int8PtrTy);
1225     if (WFn)
1226       ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy);
1227     llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(OutlinedFn, CGM.Int8PtrTy);
1228 
1229     // Create a private scope that will globalize the arguments
1230     // passed from the outside of the target region.
1231     // TODO: Is that needed?
1232     CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF);
1233 
1234     Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca(
1235         llvm::ArrayType::get(CGM.VoidPtrTy, CapturedVars.size()),
1236         "captured_vars_addrs");
1237     // There's something to share.
1238     if (!CapturedVars.empty()) {
1239       // Prepare for parallel region. Indicate the outlined function.
1240       ASTContext &Ctx = CGF.getContext();
1241       unsigned Idx = 0;
1242       for (llvm::Value *V : CapturedVars) {
1243         Address Dst = Bld.CreateConstArrayGEP(CapturedVarsAddrs, Idx);
1244         llvm::Value *PtrV;
1245         if (V->getType()->isIntegerTy())
1246           PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy);
1247         else
1248           PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy);
1249         CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false,
1250                               Ctx.getPointerType(Ctx.VoidPtrTy));
1251         ++Idx;
1252       }
1253     }
1254 
1255     llvm::Value *IfCondVal = nullptr;
1256     if (IfCond)
1257       IfCondVal = Bld.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.Int32Ty,
1258                                     /* isSigned */ false);
1259     else
1260       IfCondVal = llvm::ConstantInt::get(CGF.Int32Ty, 1);
1261 
1262     if (!NumThreadsVal)
1263       NumThreadsVal = llvm::ConstantInt::get(CGF.Int32Ty, -1);
1264     else
1265       NumThreadsVal = Bld.CreateZExtOrTrunc(NumThreadsVal, CGF.Int32Ty),
1266 
1267       assert(IfCondVal && "Expected a value");
1268     llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
1269     llvm::Value *Args[] = {
1270         RTLoc,
1271         getThreadID(CGF, Loc),
1272         IfCondVal,
1273         NumThreadsVal,
1274         llvm::ConstantInt::get(CGF.Int32Ty, -1),
1275         FnPtr,
1276         ID,
1277         Bld.CreateBitOrPointerCast(CapturedVarsAddrs.getPointer(),
1278                                    CGF.VoidPtrPtrTy),
1279         llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())};
1280     CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1281                             CGM.getModule(), OMPRTL___kmpc_parallel_51),
1282                         Args);
1283   };
1284 
1285   RegionCodeGenTy RCG(ParallelGen);
1286   RCG(CGF);
1287 }
1288 
1289 void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) {
1290   // Always emit simple barriers!
1291   if (!CGF.HaveInsertPoint())
1292     return;
1293   // Build call __kmpc_barrier_simple_spmd(nullptr, 0);
1294   // This function does not use parameters, so we can emit just default values.
1295   llvm::Value *Args[] = {
1296       llvm::ConstantPointerNull::get(
1297           cast<llvm::PointerType>(getIdentTyPointerTy())),
1298       llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)};
1299   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1300                           CGM.getModule(), OMPRTL___kmpc_barrier_simple_spmd),
1301                       Args);
1302 }
1303 
1304 void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF,
1305                                            SourceLocation Loc,
1306                                            OpenMPDirectiveKind Kind, bool,
1307                                            bool) {
1308   // Always emit simple barriers!
1309   if (!CGF.HaveInsertPoint())
1310     return;
1311   // Build call __kmpc_cancel_barrier(loc, thread_id);
1312   unsigned Flags = getDefaultFlagsForBarriers(Kind);
1313   llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags),
1314                          getThreadID(CGF, Loc)};
1315 
1316   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1317                           CGM.getModule(), OMPRTL___kmpc_barrier),
1318                       Args);
1319 }
1320 
1321 void CGOpenMPRuntimeGPU::emitCriticalRegion(
1322     CodeGenFunction &CGF, StringRef CriticalName,
1323     const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc,
1324     const Expr *Hint) {
1325   llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop");
1326   llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test");
1327   llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync");
1328   llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body");
1329   llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit");
1330 
1331   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1332 
1333   // Get the mask of active threads in the warp.
1334   llvm::Value *Mask = CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1335       CGM.getModule(), OMPRTL___kmpc_warp_active_thread_mask));
1336   // Fetch team-local id of the thread.
1337   llvm::Value *ThreadID = RT.getGPUThreadID(CGF);
1338 
1339   // Get the width of the team.
1340   llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF);
1341 
1342   // Initialize the counter variable for the loop.
1343   QualType Int32Ty =
1344       CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0);
1345   Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter");
1346   LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty);
1347   CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal,
1348                         /*isInit=*/true);
1349 
1350   // Block checks if loop counter exceeds upper bound.
1351   CGF.EmitBlock(LoopBB);
1352   llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
1353   llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth);
1354   CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB);
1355 
1356   // Block tests which single thread should execute region, and which threads
1357   // should go straight to synchronisation point.
1358   CGF.EmitBlock(TestBB);
1359   CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
1360   llvm::Value *CmpThreadToCounter =
1361       CGF.Builder.CreateICmpEQ(ThreadID, CounterVal);
1362   CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB);
1363 
1364   // Block emits the body of the critical region.
1365   CGF.EmitBlock(BodyBB);
1366 
1367   // Output the critical statement.
1368   CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc,
1369                                       Hint);
1370 
1371   // After the body surrounded by the critical region, the single executing
1372   // thread will jump to the synchronisation point.
1373   // Block waits for all threads in current team to finish then increments the
1374   // counter variable and returns to the loop.
1375   CGF.EmitBlock(SyncBB);
1376   // Reconverge active threads in the warp.
1377   (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1378                                 CGM.getModule(), OMPRTL___kmpc_syncwarp),
1379                             Mask);
1380 
1381   llvm::Value *IncCounterVal =
1382       CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1));
1383   CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal);
1384   CGF.EmitBranch(LoopBB);
1385 
1386   // Block that is reached when  all threads in the team complete the region.
1387   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1388 }
1389 
1390 /// Cast value to the specified type.
1391 static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val,
1392                                     QualType ValTy, QualType CastTy,
1393                                     SourceLocation Loc) {
1394   assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() &&
1395          "Cast type must sized.");
1396   assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() &&
1397          "Val type must sized.");
1398   llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy);
1399   if (ValTy == CastTy)
1400     return Val;
1401   if (CGF.getContext().getTypeSizeInChars(ValTy) ==
1402       CGF.getContext().getTypeSizeInChars(CastTy))
1403     return CGF.Builder.CreateBitCast(Val, LLVMCastTy);
1404   if (CastTy->isIntegerType() && ValTy->isIntegerType())
1405     return CGF.Builder.CreateIntCast(Val, LLVMCastTy,
1406                                      CastTy->hasSignedIntegerRepresentation());
1407   Address CastItem = CGF.CreateMemTemp(CastTy);
1408   Address ValCastItem = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
1409       CastItem, Val->getType()->getPointerTo(CastItem.getAddressSpace()),
1410       Val->getType());
1411   CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy,
1412                         LValueBaseInfo(AlignmentSource::Type),
1413                         TBAAAccessInfo());
1414   return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc,
1415                               LValueBaseInfo(AlignmentSource::Type),
1416                               TBAAAccessInfo());
1417 }
1418 
1419 /// This function creates calls to one of two shuffle functions to copy
1420 /// variables between lanes in a warp.
1421 static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF,
1422                                                  llvm::Value *Elem,
1423                                                  QualType ElemType,
1424                                                  llvm::Value *Offset,
1425                                                  SourceLocation Loc) {
1426   CodeGenModule &CGM = CGF.CGM;
1427   CGBuilderTy &Bld = CGF.Builder;
1428   CGOpenMPRuntimeGPU &RT =
1429       *(static_cast<CGOpenMPRuntimeGPU *>(&CGM.getOpenMPRuntime()));
1430   llvm::OpenMPIRBuilder &OMPBuilder = RT.getOMPBuilder();
1431 
1432   CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType);
1433   assert(Size.getQuantity() <= 8 &&
1434          "Unsupported bitwidth in shuffle instruction.");
1435 
1436   RuntimeFunction ShuffleFn = Size.getQuantity() <= 4
1437                                   ? OMPRTL___kmpc_shuffle_int32
1438                                   : OMPRTL___kmpc_shuffle_int64;
1439 
1440   // Cast all types to 32- or 64-bit values before calling shuffle routines.
1441   QualType CastTy = CGF.getContext().getIntTypeForBitwidth(
1442       Size.getQuantity() <= 4 ? 32 : 64, /*Signed=*/1);
1443   llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc);
1444   llvm::Value *WarpSize =
1445       Bld.CreateIntCast(RT.getGPUWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true);
1446 
1447   llvm::Value *ShuffledVal = CGF.EmitRuntimeCall(
1448       OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), ShuffleFn),
1449       {ElemCast, Offset, WarpSize});
1450 
1451   return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc);
1452 }
1453 
1454 static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr,
1455                             Address DestAddr, QualType ElemType,
1456                             llvm::Value *Offset, SourceLocation Loc) {
1457   CGBuilderTy &Bld = CGF.Builder;
1458 
1459   CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType);
1460   // Create the loop over the big sized data.
1461   // ptr = (void*)Elem;
1462   // ptrEnd = (void*) Elem + 1;
1463   // Step = 8;
1464   // while (ptr + Step < ptrEnd)
1465   //   shuffle((int64_t)*ptr);
1466   // Step = 4;
1467   // while (ptr + Step < ptrEnd)
1468   //   shuffle((int32_t)*ptr);
1469   // ...
1470   Address ElemPtr = DestAddr;
1471   Address Ptr = SrcAddr;
1472   Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast(
1473       Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy, CGF.Int8Ty);
1474   for (int IntSize = 8; IntSize >= 1; IntSize /= 2) {
1475     if (Size < CharUnits::fromQuantity(IntSize))
1476       continue;
1477     QualType IntType = CGF.getContext().getIntTypeForBitwidth(
1478         CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)),
1479         /*Signed=*/1);
1480     llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType);
1481     Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo(),
1482                                                   IntTy);
1483     ElemPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
1484         ElemPtr, IntTy->getPointerTo(), IntTy);
1485     if (Size.getQuantity() / IntSize > 1) {
1486       llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond");
1487       llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then");
1488       llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit");
1489       llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock();
1490       CGF.EmitBlock(PreCondBB);
1491       llvm::PHINode *PhiSrc =
1492           Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2);
1493       PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB);
1494       llvm::PHINode *PhiDest =
1495           Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2);
1496       PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB);
1497       Ptr = Address(PhiSrc, Ptr.getElementType(), Ptr.getAlignment());
1498       ElemPtr =
1499           Address(PhiDest, ElemPtr.getElementType(), ElemPtr.getAlignment());
1500       llvm::Value *PtrDiff = Bld.CreatePtrDiff(
1501           CGF.Int8Ty, PtrEnd.getPointer(),
1502           Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr.getPointer(),
1503                                                   CGF.VoidPtrTy));
1504       Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)),
1505                        ThenBB, ExitBB);
1506       CGF.EmitBlock(ThenBB);
1507       llvm::Value *Res = createRuntimeShuffleFunction(
1508           CGF,
1509           CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc,
1510                                LValueBaseInfo(AlignmentSource::Type),
1511                                TBAAAccessInfo()),
1512           IntType, Offset, Loc);
1513       CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType,
1514                             LValueBaseInfo(AlignmentSource::Type),
1515                             TBAAAccessInfo());
1516       Address LocalPtr = Bld.CreateConstGEP(Ptr, 1);
1517       Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1);
1518       PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB);
1519       PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB);
1520       CGF.EmitBranch(PreCondBB);
1521       CGF.EmitBlock(ExitBB);
1522     } else {
1523       llvm::Value *Res = createRuntimeShuffleFunction(
1524           CGF,
1525           CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc,
1526                                LValueBaseInfo(AlignmentSource::Type),
1527                                TBAAAccessInfo()),
1528           IntType, Offset, Loc);
1529       CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType,
1530                             LValueBaseInfo(AlignmentSource::Type),
1531                             TBAAAccessInfo());
1532       Ptr = Bld.CreateConstGEP(Ptr, 1);
1533       ElemPtr = Bld.CreateConstGEP(ElemPtr, 1);
1534     }
1535     Size = Size % IntSize;
1536   }
1537 }
1538 
1539 namespace {
1540 enum CopyAction : unsigned {
1541   // RemoteLaneToThread: Copy over a Reduce list from a remote lane in
1542   // the warp using shuffle instructions.
1543   RemoteLaneToThread,
1544   // ThreadCopy: Make a copy of a Reduce list on the thread's stack.
1545   ThreadCopy,
1546   // ThreadToScratchpad: Copy a team-reduced array to the scratchpad.
1547   ThreadToScratchpad,
1548   // ScratchpadToThread: Copy from a scratchpad array in global memory
1549   // containing team-reduced data to a thread's stack.
1550   ScratchpadToThread,
1551 };
1552 } // namespace
1553 
1554 struct CopyOptionsTy {
1555   llvm::Value *RemoteLaneOffset;
1556   llvm::Value *ScratchpadIndex;
1557   llvm::Value *ScratchpadWidth;
1558 };
1559 
1560 /// Emit instructions to copy a Reduce list, which contains partially
1561 /// aggregated values, in the specified direction.
1562 static void emitReductionListCopy(
1563     CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy,
1564     ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase,
1565     CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) {
1566 
1567   CodeGenModule &CGM = CGF.CGM;
1568   ASTContext &C = CGM.getContext();
1569   CGBuilderTy &Bld = CGF.Builder;
1570 
1571   llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset;
1572   llvm::Value *ScratchpadIndex = CopyOptions.ScratchpadIndex;
1573   llvm::Value *ScratchpadWidth = CopyOptions.ScratchpadWidth;
1574 
1575   // Iterates, element-by-element, through the source Reduce list and
1576   // make a copy.
1577   unsigned Idx = 0;
1578   unsigned Size = Privates.size();
1579   for (const Expr *Private : Privates) {
1580     Address SrcElementAddr = Address::invalid();
1581     Address DestElementAddr = Address::invalid();
1582     Address DestElementPtrAddr = Address::invalid();
1583     // Should we shuffle in an element from a remote lane?
1584     bool ShuffleInElement = false;
1585     // Set to true to update the pointer in the dest Reduce list to a
1586     // newly created element.
1587     bool UpdateDestListPtr = false;
1588     // Increment the src or dest pointer to the scratchpad, for each
1589     // new element.
1590     bool IncrScratchpadSrc = false;
1591     bool IncrScratchpadDest = false;
1592     QualType PrivatePtrType = C.getPointerType(Private->getType());
1593     llvm::Type *PrivateLlvmPtrType = CGF.ConvertType(PrivatePtrType);
1594 
1595     switch (Action) {
1596     case RemoteLaneToThread: {
1597       // Step 1.1: Get the address for the src element in the Reduce list.
1598       Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
1599       SrcElementAddr = CGF.EmitLoadOfPointer(
1600           SrcElementPtrAddr.withElementType(PrivateLlvmPtrType),
1601           PrivatePtrType->castAs<PointerType>());
1602 
1603       // Step 1.2: Create a temporary to store the element in the destination
1604       // Reduce list.
1605       DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
1606       DestElementAddr =
1607           CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element");
1608       ShuffleInElement = true;
1609       UpdateDestListPtr = true;
1610       break;
1611     }
1612     case ThreadCopy: {
1613       // Step 1.1: Get the address for the src element in the Reduce list.
1614       Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
1615       SrcElementAddr = CGF.EmitLoadOfPointer(
1616           SrcElementPtrAddr.withElementType(PrivateLlvmPtrType),
1617           PrivatePtrType->castAs<PointerType>());
1618 
1619       // Step 1.2: Get the address for dest element.  The destination
1620       // element has already been created on the thread's stack.
1621       DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
1622       DestElementAddr = CGF.EmitLoadOfPointer(
1623           DestElementPtrAddr.withElementType(PrivateLlvmPtrType),
1624           PrivatePtrType->castAs<PointerType>());
1625       break;
1626     }
1627     case ThreadToScratchpad: {
1628       // Step 1.1: Get the address for the src element in the Reduce list.
1629       Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
1630       SrcElementAddr = CGF.EmitLoadOfPointer(
1631           SrcElementPtrAddr.withElementType(PrivateLlvmPtrType),
1632           PrivatePtrType->castAs<PointerType>());
1633 
1634       // Step 1.2: Get the address for dest element:
1635       // address = base + index * ElementSizeInChars.
1636       llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
1637       llvm::Value *CurrentOffset =
1638           Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex);
1639       llvm::Value *ScratchPadElemAbsolutePtrVal =
1640           Bld.CreateNUWAdd(DestBase.getPointer(), CurrentOffset);
1641       ScratchPadElemAbsolutePtrVal =
1642           Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy);
1643       DestElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty,
1644                                 C.getTypeAlignInChars(Private->getType()));
1645       IncrScratchpadDest = true;
1646       break;
1647     }
1648     case ScratchpadToThread: {
1649       // Step 1.1: Get the address for the src element in the scratchpad.
1650       // address = base + index * ElementSizeInChars.
1651       llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
1652       llvm::Value *CurrentOffset =
1653           Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex);
1654       llvm::Value *ScratchPadElemAbsolutePtrVal =
1655           Bld.CreateNUWAdd(SrcBase.getPointer(), CurrentOffset);
1656       ScratchPadElemAbsolutePtrVal =
1657           Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy);
1658       SrcElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty,
1659                                C.getTypeAlignInChars(Private->getType()));
1660       IncrScratchpadSrc = true;
1661 
1662       // Step 1.2: Create a temporary to store the element in the destination
1663       // Reduce list.
1664       DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
1665       DestElementAddr =
1666           CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element");
1667       UpdateDestListPtr = true;
1668       break;
1669     }
1670     }
1671 
1672     // Regardless of src and dest of copy, we emit the load of src
1673     // element as this is required in all directions
1674     SrcElementAddr = SrcElementAddr.withElementType(
1675         CGF.ConvertTypeForMem(Private->getType()));
1676     DestElementAddr =
1677         DestElementAddr.withElementType(SrcElementAddr.getElementType());
1678 
1679     // Now that all active lanes have read the element in the
1680     // Reduce list, shuffle over the value from the remote lane.
1681     if (ShuffleInElement) {
1682       shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(),
1683                       RemoteLaneOffset, Private->getExprLoc());
1684     } else {
1685       switch (CGF.getEvaluationKind(Private->getType())) {
1686       case TEK_Scalar: {
1687         llvm::Value *Elem = CGF.EmitLoadOfScalar(
1688             SrcElementAddr, /*Volatile=*/false, Private->getType(),
1689             Private->getExprLoc(), LValueBaseInfo(AlignmentSource::Type),
1690             TBAAAccessInfo());
1691         // Store the source element value to the dest element address.
1692         CGF.EmitStoreOfScalar(
1693             Elem, DestElementAddr, /*Volatile=*/false, Private->getType(),
1694             LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
1695         break;
1696       }
1697       case TEK_Complex: {
1698         CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex(
1699             CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
1700             Private->getExprLoc());
1701         CGF.EmitStoreOfComplex(
1702             Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
1703             /*isInit=*/false);
1704         break;
1705       }
1706       case TEK_Aggregate:
1707         CGF.EmitAggregateCopy(
1708             CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
1709             CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
1710             Private->getType(), AggValueSlot::DoesNotOverlap);
1711         break;
1712       }
1713     }
1714 
1715     // Step 3.1: Modify reference in dest Reduce list as needed.
1716     // Modifying the reference in Reduce list to point to the newly
1717     // created element.  The element is live in the current function
1718     // scope and that of functions it invokes (i.e., reduce_function).
1719     // RemoteReduceData[i] = (void*)&RemoteElem
1720     if (UpdateDestListPtr) {
1721       CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast(
1722                                 DestElementAddr.getPointer(), CGF.VoidPtrTy),
1723                             DestElementPtrAddr, /*Volatile=*/false,
1724                             C.VoidPtrTy);
1725     }
1726 
1727     // Step 4.1: Increment SrcBase/DestBase so that it points to the starting
1728     // address of the next element in scratchpad memory, unless we're currently
1729     // processing the last one.  Memory alignment is also taken care of here.
1730     if ((IncrScratchpadDest || IncrScratchpadSrc) && (Idx + 1 < Size)) {
1731       // FIXME: This code doesn't make any sense, it's trying to perform
1732       // integer arithmetic on pointers.
1733       llvm::Value *ScratchpadBasePtr =
1734           IncrScratchpadDest ? DestBase.getPointer() : SrcBase.getPointer();
1735       llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
1736       ScratchpadBasePtr = Bld.CreateNUWAdd(
1737           ScratchpadBasePtr,
1738           Bld.CreateNUWMul(ScratchpadWidth, ElementSizeInChars));
1739 
1740       // Take care of global memory alignment for performance
1741       ScratchpadBasePtr = Bld.CreateNUWSub(
1742           ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1));
1743       ScratchpadBasePtr = Bld.CreateUDiv(
1744           ScratchpadBasePtr,
1745           llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment));
1746       ScratchpadBasePtr = Bld.CreateNUWAdd(
1747           ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1));
1748       ScratchpadBasePtr = Bld.CreateNUWMul(
1749           ScratchpadBasePtr,
1750           llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment));
1751 
1752       if (IncrScratchpadDest)
1753         DestBase =
1754             Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign());
1755       else /* IncrScratchpadSrc = true */
1756         SrcBase =
1757             Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign());
1758     }
1759 
1760     ++Idx;
1761   }
1762 }
1763 
1764 /// This function emits a helper that gathers Reduce lists from the first
1765 /// lane of every active warp to lanes in the first warp.
1766 ///
1767 /// void inter_warp_copy_func(void* reduce_data, num_warps)
1768 ///   shared smem[warp_size];
1769 ///   For all data entries D in reduce_data:
1770 ///     sync
1771 ///     If (I am the first lane in each warp)
1772 ///       Copy my local D to smem[warp_id]
1773 ///     sync
1774 ///     if (I am the first warp)
1775 ///       Copy smem[thread_id] to my local D
1776 static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM,
1777                                               ArrayRef<const Expr *> Privates,
1778                                               QualType ReductionArrayTy,
1779                                               SourceLocation Loc) {
1780   ASTContext &C = CGM.getContext();
1781   llvm::Module &M = CGM.getModule();
1782 
1783   // ReduceList: thread local Reduce list.
1784   // At the stage of the computation when this function is called, partially
1785   // aggregated values reside in the first lane of every active warp.
1786   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
1787                                   C.VoidPtrTy, ImplicitParamDecl::Other);
1788   // NumWarps: number of warps active in the parallel region.  This could
1789   // be smaller than 32 (max warps in a CTA) for partial block reduction.
1790   ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
1791                                 C.getIntTypeForBitwidth(32, /* Signed */ true),
1792                                 ImplicitParamDecl::Other);
1793   FunctionArgList Args;
1794   Args.push_back(&ReduceListArg);
1795   Args.push_back(&NumWarpsArg);
1796 
1797   const CGFunctionInfo &CGFI =
1798       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
1799   auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI),
1800                                     llvm::GlobalValue::InternalLinkage,
1801                                     "_omp_reduction_inter_warp_copy_func", &M);
1802   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
1803   Fn->setDoesNotRecurse();
1804   CodeGenFunction CGF(CGM);
1805   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
1806 
1807   CGBuilderTy &Bld = CGF.Builder;
1808 
1809   // This array is used as a medium to transfer, one reduce element at a time,
1810   // the data from the first lane of every warp to lanes in the first warp
1811   // in order to perform the final step of a reduction in a parallel region
1812   // (reduction across warps).  The array is placed in NVPTX __shared__ memory
1813   // for reduced latency, as well as to have a distinct copy for concurrently
1814   // executing target regions.  The array is declared with common linkage so
1815   // as to be shared across compilation units.
1816   StringRef TransferMediumName =
1817       "__openmp_nvptx_data_transfer_temporary_storage";
1818   llvm::GlobalVariable *TransferMedium =
1819       M.getGlobalVariable(TransferMediumName);
1820   unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size;
1821   if (!TransferMedium) {
1822     auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize);
1823     unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared);
1824     TransferMedium = new llvm::GlobalVariable(
1825         M, Ty, /*isConstant=*/false, llvm::GlobalVariable::WeakAnyLinkage,
1826         llvm::UndefValue::get(Ty), TransferMediumName,
1827         /*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal,
1828         SharedAddressSpace);
1829     CGM.addCompilerUsedGlobal(TransferMedium);
1830   }
1831 
1832   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1833   // Get the CUDA thread id of the current OpenMP thread on the GPU.
1834   llvm::Value *ThreadID = RT.getGPUThreadID(CGF);
1835   // nvptx_lane_id = nvptx_id % warpsize
1836   llvm::Value *LaneID = getNVPTXLaneID(CGF);
1837   // nvptx_warp_id = nvptx_id / warpsize
1838   llvm::Value *WarpID = getNVPTXWarpID(CGF);
1839 
1840   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
1841   llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
1842   Address LocalReduceList(
1843       Bld.CreatePointerBitCastOrAddrSpaceCast(
1844           CGF.EmitLoadOfScalar(
1845               AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc,
1846               LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()),
1847           ElemTy->getPointerTo()),
1848       ElemTy, CGF.getPointerAlign());
1849 
1850   unsigned Idx = 0;
1851   for (const Expr *Private : Privates) {
1852     //
1853     // Warp master copies reduce element to transfer medium in __shared__
1854     // memory.
1855     //
1856     unsigned RealTySize =
1857         C.getTypeSizeInChars(Private->getType())
1858             .alignTo(C.getTypeAlignInChars(Private->getType()))
1859             .getQuantity();
1860     for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /=2) {
1861       unsigned NumIters = RealTySize / TySize;
1862       if (NumIters == 0)
1863         continue;
1864       QualType CType = C.getIntTypeForBitwidth(
1865           C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1);
1866       llvm::Type *CopyType = CGF.ConvertTypeForMem(CType);
1867       CharUnits Align = CharUnits::fromQuantity(TySize);
1868       llvm::Value *Cnt = nullptr;
1869       Address CntAddr = Address::invalid();
1870       llvm::BasicBlock *PrecondBB = nullptr;
1871       llvm::BasicBlock *ExitBB = nullptr;
1872       if (NumIters > 1) {
1873         CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr");
1874         CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr,
1875                               /*Volatile=*/false, C.IntTy);
1876         PrecondBB = CGF.createBasicBlock("precond");
1877         ExitBB = CGF.createBasicBlock("exit");
1878         llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body");
1879         // There is no need to emit line number for unconditional branch.
1880         (void)ApplyDebugLocation::CreateEmpty(CGF);
1881         CGF.EmitBlock(PrecondBB);
1882         Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc);
1883         llvm::Value *Cmp =
1884             Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters));
1885         Bld.CreateCondBr(Cmp, BodyBB, ExitBB);
1886         CGF.EmitBlock(BodyBB);
1887       }
1888       // kmpc_barrier.
1889       CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
1890                                              /*EmitChecks=*/false,
1891                                              /*ForceSimpleCall=*/true);
1892       llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then");
1893       llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else");
1894       llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont");
1895 
1896       // if (lane_id == 0)
1897       llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master");
1898       Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB);
1899       CGF.EmitBlock(ThenBB);
1900 
1901       // Reduce element = LocalReduceList[i]
1902       Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
1903       llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
1904           ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
1905       // elemptr = ((CopyType*)(elemptrptr)) + I
1906       Address ElemPtr(ElemPtrPtr, CopyType, Align);
1907       if (NumIters > 1)
1908         ElemPtr = Bld.CreateGEP(ElemPtr, Cnt);
1909 
1910       // Get pointer to location in transfer medium.
1911       // MediumPtr = &medium[warp_id]
1912       llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP(
1913           TransferMedium->getValueType(), TransferMedium,
1914           {llvm::Constant::getNullValue(CGM.Int64Ty), WarpID});
1915       // Casting to actual data type.
1916       // MediumPtr = (CopyType*)MediumPtrAddr;
1917       Address MediumPtr(
1918           Bld.CreateBitCast(
1919               MediumPtrVal,
1920               CopyType->getPointerTo(
1921                   MediumPtrVal->getType()->getPointerAddressSpace())),
1922           CopyType, Align);
1923 
1924       // elem = *elemptr
1925       //*MediumPtr = elem
1926       llvm::Value *Elem = CGF.EmitLoadOfScalar(
1927           ElemPtr, /*Volatile=*/false, CType, Loc,
1928           LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
1929       // Store the source element value to the dest element address.
1930       CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType,
1931                             LValueBaseInfo(AlignmentSource::Type),
1932                             TBAAAccessInfo());
1933 
1934       Bld.CreateBr(MergeBB);
1935 
1936       CGF.EmitBlock(ElseBB);
1937       Bld.CreateBr(MergeBB);
1938 
1939       CGF.EmitBlock(MergeBB);
1940 
1941       // kmpc_barrier.
1942       CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
1943                                              /*EmitChecks=*/false,
1944                                              /*ForceSimpleCall=*/true);
1945 
1946       //
1947       // Warp 0 copies reduce element from transfer medium.
1948       //
1949       llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then");
1950       llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else");
1951       llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont");
1952 
1953       Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg);
1954       llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar(
1955           AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc);
1956 
1957       // Up to 32 threads in warp 0 are active.
1958       llvm::Value *IsActiveThread =
1959           Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread");
1960       Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB);
1961 
1962       CGF.EmitBlock(W0ThenBB);
1963 
1964       // SrcMediumPtr = &medium[tid]
1965       llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP(
1966           TransferMedium->getValueType(), TransferMedium,
1967           {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID});
1968       // SrcMediumVal = *SrcMediumPtr;
1969       Address SrcMediumPtr(
1970           Bld.CreateBitCast(
1971               SrcMediumPtrVal,
1972               CopyType->getPointerTo(
1973                   SrcMediumPtrVal->getType()->getPointerAddressSpace())),
1974           CopyType, Align);
1975 
1976       // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I
1977       Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
1978       llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar(
1979           TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc);
1980       Address TargetElemPtr(TargetElemPtrVal, CopyType, Align);
1981       if (NumIters > 1)
1982         TargetElemPtr = Bld.CreateGEP(TargetElemPtr, Cnt);
1983 
1984       // *TargetElemPtr = SrcMediumVal;
1985       llvm::Value *SrcMediumValue =
1986           CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc);
1987       CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false,
1988                             CType);
1989       Bld.CreateBr(W0MergeBB);
1990 
1991       CGF.EmitBlock(W0ElseBB);
1992       Bld.CreateBr(W0MergeBB);
1993 
1994       CGF.EmitBlock(W0MergeBB);
1995 
1996       if (NumIters > 1) {
1997         Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1));
1998         CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy);
1999         CGF.EmitBranch(PrecondBB);
2000         (void)ApplyDebugLocation::CreateEmpty(CGF);
2001         CGF.EmitBlock(ExitBB);
2002       }
2003       RealTySize %= TySize;
2004     }
2005     ++Idx;
2006   }
2007 
2008   CGF.FinishFunction();
2009   return Fn;
2010 }
2011 
2012 /// Emit a helper that reduces data across two OpenMP threads (lanes)
2013 /// in the same warp.  It uses shuffle instructions to copy over data from
2014 /// a remote lane's stack.  The reduction algorithm performed is specified
2015 /// by the fourth parameter.
2016 ///
2017 /// Algorithm Versions.
2018 /// Full Warp Reduce (argument value 0):
2019 ///   This algorithm assumes that all 32 lanes are active and gathers
2020 ///   data from these 32 lanes, producing a single resultant value.
2021 /// Contiguous Partial Warp Reduce (argument value 1):
2022 ///   This algorithm assumes that only a *contiguous* subset of lanes
2023 ///   are active.  This happens for the last warp in a parallel region
2024 ///   when the user specified num_threads is not an integer multiple of
2025 ///   32.  This contiguous subset always starts with the zeroth lane.
2026 /// Partial Warp Reduce (argument value 2):
2027 ///   This algorithm gathers data from any number of lanes at any position.
2028 /// All reduced values are stored in the lowest possible lane.  The set
2029 /// of problems every algorithm addresses is a super set of those
2030 /// addressable by algorithms with a lower version number.  Overhead
2031 /// increases as algorithm version increases.
2032 ///
2033 /// Terminology
2034 /// Reduce element:
2035 ///   Reduce element refers to the individual data field with primitive
2036 ///   data types to be combined and reduced across threads.
2037 /// Reduce list:
2038 ///   Reduce list refers to a collection of local, thread-private
2039 ///   reduce elements.
2040 /// Remote Reduce list:
2041 ///   Remote Reduce list refers to a collection of remote (relative to
2042 ///   the current thread) reduce elements.
2043 ///
2044 /// We distinguish between three states of threads that are important to
2045 /// the implementation of this function.
2046 /// Alive threads:
2047 ///   Threads in a warp executing the SIMT instruction, as distinguished from
2048 ///   threads that are inactive due to divergent control flow.
2049 /// Active threads:
2050 ///   The minimal set of threads that has to be alive upon entry to this
2051 ///   function.  The computation is correct iff active threads are alive.
2052 ///   Some threads are alive but they are not active because they do not
2053 ///   contribute to the computation in any useful manner.  Turning them off
2054 ///   may introduce control flow overheads without any tangible benefits.
2055 /// Effective threads:
2056 ///   In order to comply with the argument requirements of the shuffle
2057 ///   function, we must keep all lanes holding data alive.  But at most
2058 ///   half of them perform value aggregation; we refer to this half of
2059 ///   threads as effective. The other half is simply handing off their
2060 ///   data.
2061 ///
2062 /// Procedure
2063 /// Value shuffle:
2064 ///   In this step active threads transfer data from higher lane positions
2065 ///   in the warp to lower lane positions, creating Remote Reduce list.
2066 /// Value aggregation:
2067 ///   In this step, effective threads combine their thread local Reduce list
2068 ///   with Remote Reduce list and store the result in the thread local
2069 ///   Reduce list.
2070 /// Value copy:
2071 ///   In this step, we deal with the assumption made by algorithm 2
2072 ///   (i.e. contiguity assumption).  When we have an odd number of lanes
2073 ///   active, say 2k+1, only k threads will be effective and therefore k
2074 ///   new values will be produced.  However, the Reduce list owned by the
2075 ///   (2k+1)th thread is ignored in the value aggregation.  Therefore
2076 ///   we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so
2077 ///   that the contiguity assumption still holds.
2078 static llvm::Function *emitShuffleAndReduceFunction(
2079     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2080     QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) {
2081   ASTContext &C = CGM.getContext();
2082 
2083   // Thread local Reduce list used to host the values of data to be reduced.
2084   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2085                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2086   // Current lane id; could be logical.
2087   ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy,
2088                               ImplicitParamDecl::Other);
2089   // Offset of the remote source lane relative to the current lane.
2090   ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2091                                         C.ShortTy, ImplicitParamDecl::Other);
2092   // Algorithm version.  This is expected to be known at compile time.
2093   ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2094                                C.ShortTy, ImplicitParamDecl::Other);
2095   FunctionArgList Args;
2096   Args.push_back(&ReduceListArg);
2097   Args.push_back(&LaneIDArg);
2098   Args.push_back(&RemoteLaneOffsetArg);
2099   Args.push_back(&AlgoVerArg);
2100 
2101   const CGFunctionInfo &CGFI =
2102       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2103   auto *Fn = llvm::Function::Create(
2104       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2105       "_omp_reduction_shuffle_and_reduce_func", &CGM.getModule());
2106   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2107   Fn->setDoesNotRecurse();
2108 
2109   CodeGenFunction CGF(CGM);
2110   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2111 
2112   CGBuilderTy &Bld = CGF.Builder;
2113 
2114   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2115   llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
2116   Address LocalReduceList(
2117       Bld.CreatePointerBitCastOrAddrSpaceCast(
2118           CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
2119                                C.VoidPtrTy, SourceLocation()),
2120           ElemTy->getPointerTo()),
2121       ElemTy, CGF.getPointerAlign());
2122 
2123   Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg);
2124   llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar(
2125       AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
2126 
2127   Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg);
2128   llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar(
2129       AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
2130 
2131   Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg);
2132   llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar(
2133       AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
2134 
2135   // Create a local thread-private variable to host the Reduce list
2136   // from a remote lane.
2137   Address RemoteReduceList =
2138       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list");
2139 
2140   // This loop iterates through the list of reduce elements and copies,
2141   // element by element, from a remote lane in the warp to RemoteReduceList,
2142   // hosted on the thread's stack.
2143   emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates,
2144                         LocalReduceList, RemoteReduceList,
2145                         {/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal,
2146                          /*ScratchpadIndex=*/nullptr,
2147                          /*ScratchpadWidth=*/nullptr});
2148 
2149   // The actions to be performed on the Remote Reduce list is dependent
2150   // on the algorithm version.
2151   //
2152   //  if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 &&
2153   //  LaneId % 2 == 0 && Offset > 0):
2154   //    do the reduction value aggregation
2155   //
2156   //  The thread local variable Reduce list is mutated in place to host the
2157   //  reduced data, which is the aggregated value produced from local and
2158   //  remote lanes.
2159   //
2160   //  Note that AlgoVer is expected to be a constant integer known at compile
2161   //  time.
2162   //  When AlgoVer==0, the first conjunction evaluates to true, making
2163   //    the entire predicate true during compile time.
2164   //  When AlgoVer==1, the second conjunction has only the second part to be
2165   //    evaluated during runtime.  Other conjunctions evaluates to false
2166   //    during compile time.
2167   //  When AlgoVer==2, the third conjunction has only the second part to be
2168   //    evaluated during runtime.  Other conjunctions evaluates to false
2169   //    during compile time.
2170   llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal);
2171 
2172   llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1));
2173   llvm::Value *CondAlgo1 = Bld.CreateAnd(
2174       Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal));
2175 
2176   llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2));
2177   llvm::Value *CondAlgo2 = Bld.CreateAnd(
2178       Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1))));
2179   CondAlgo2 = Bld.CreateAnd(
2180       CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0)));
2181 
2182   llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1);
2183   CondReduce = Bld.CreateOr(CondReduce, CondAlgo2);
2184 
2185   llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then");
2186   llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else");
2187   llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont");
2188   Bld.CreateCondBr(CondReduce, ThenBB, ElseBB);
2189 
2190   CGF.EmitBlock(ThenBB);
2191   // reduce_function(LocalReduceList, RemoteReduceList)
2192   llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2193       LocalReduceList.getPointer(), CGF.VoidPtrTy);
2194   llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2195       RemoteReduceList.getPointer(), CGF.VoidPtrTy);
2196   CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
2197       CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr});
2198   Bld.CreateBr(MergeBB);
2199 
2200   CGF.EmitBlock(ElseBB);
2201   Bld.CreateBr(MergeBB);
2202 
2203   CGF.EmitBlock(MergeBB);
2204 
2205   // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local
2206   // Reduce list.
2207   Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1));
2208   llvm::Value *CondCopy = Bld.CreateAnd(
2209       Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal));
2210 
2211   llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then");
2212   llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else");
2213   llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont");
2214   Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB);
2215 
2216   CGF.EmitBlock(CpyThenBB);
2217   emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates,
2218                         RemoteReduceList, LocalReduceList);
2219   Bld.CreateBr(CpyMergeBB);
2220 
2221   CGF.EmitBlock(CpyElseBB);
2222   Bld.CreateBr(CpyMergeBB);
2223 
2224   CGF.EmitBlock(CpyMergeBB);
2225 
2226   CGF.FinishFunction();
2227   return Fn;
2228 }
2229 
2230 /// This function emits a helper that copies all the reduction variables from
2231 /// the team into the provided global buffer for the reduction variables.
2232 ///
2233 /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
2234 ///   For all data entries D in reduce_data:
2235 ///     Copy local D to buffer.D[Idx]
2236 static llvm::Value *emitListToGlobalCopyFunction(
2237     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2238     QualType ReductionArrayTy, SourceLocation Loc,
2239     const RecordDecl *TeamReductionRec,
2240     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2241         &VarFieldMap) {
2242   ASTContext &C = CGM.getContext();
2243 
2244   // Buffer: global reduction buffer.
2245   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2246                               C.VoidPtrTy, ImplicitParamDecl::Other);
2247   // Idx: index of the buffer.
2248   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2249                            ImplicitParamDecl::Other);
2250   // ReduceList: thread local Reduce list.
2251   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2252                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2253   FunctionArgList Args;
2254   Args.push_back(&BufferArg);
2255   Args.push_back(&IdxArg);
2256   Args.push_back(&ReduceListArg);
2257 
2258   const CGFunctionInfo &CGFI =
2259       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2260   auto *Fn = llvm::Function::Create(
2261       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2262       "_omp_reduction_list_to_global_copy_func", &CGM.getModule());
2263   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2264   Fn->setDoesNotRecurse();
2265   CodeGenFunction CGF(CGM);
2266   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2267 
2268   CGBuilderTy &Bld = CGF.Builder;
2269 
2270   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2271   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2272   llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
2273   Address LocalReduceList(
2274       Bld.CreatePointerBitCastOrAddrSpaceCast(
2275           CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
2276                                C.VoidPtrTy, Loc),
2277           ElemTy->getPointerTo()),
2278       ElemTy, CGF.getPointerAlign());
2279   QualType StaticTy = C.getRecordType(TeamReductionRec);
2280   llvm::Type *LLVMReductionsBufferTy =
2281       CGM.getTypes().ConvertTypeForMem(StaticTy);
2282   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2283       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2284       LLVMReductionsBufferTy->getPointerTo());
2285   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2286                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2287                                               /*Volatile=*/false, C.IntTy,
2288                                               Loc)};
2289   unsigned Idx = 0;
2290   for (const Expr *Private : Privates) {
2291     // Reduce element = LocalReduceList[i]
2292     Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
2293     llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
2294         ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
2295     // elemptr = ((CopyType*)(elemptrptr)) + I
2296     ElemTy = CGF.ConvertTypeForMem(Private->getType());
2297     ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2298         ElemPtrPtr, ElemTy->getPointerTo());
2299     Address ElemPtr =
2300         Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType()));
2301     const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl();
2302     // Global = Buffer.VD[Idx];
2303     const FieldDecl *FD = VarFieldMap.lookup(VD);
2304     LValue GlobLVal = CGF.EmitLValueForField(
2305         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2306     Address GlobAddr = GlobLVal.getAddress(CGF);
2307     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(),
2308                                                    GlobAddr.getPointer(), Idxs);
2309     GlobLVal.setAddress(Address(BufferPtr,
2310                                 CGF.ConvertTypeForMem(Private->getType()),
2311                                 GlobAddr.getAlignment()));
2312     switch (CGF.getEvaluationKind(Private->getType())) {
2313     case TEK_Scalar: {
2314       llvm::Value *V = CGF.EmitLoadOfScalar(
2315           ElemPtr, /*Volatile=*/false, Private->getType(), Loc,
2316           LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
2317       CGF.EmitStoreOfScalar(V, GlobLVal);
2318       break;
2319     }
2320     case TEK_Complex: {
2321       CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(
2322           CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc);
2323       CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false);
2324       break;
2325     }
2326     case TEK_Aggregate:
2327       CGF.EmitAggregateCopy(GlobLVal,
2328                             CGF.MakeAddrLValue(ElemPtr, Private->getType()),
2329                             Private->getType(), AggValueSlot::DoesNotOverlap);
2330       break;
2331     }
2332     ++Idx;
2333   }
2334 
2335   CGF.FinishFunction();
2336   return Fn;
2337 }
2338 
2339 /// This function emits a helper that reduces all the reduction variables from
2340 /// the team into the provided global buffer for the reduction variables.
2341 ///
2342 /// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data)
2343 ///  void *GlobPtrs[];
2344 ///  GlobPtrs[0] = (void*)&buffer.D0[Idx];
2345 ///  ...
2346 ///  GlobPtrs[N] = (void*)&buffer.DN[Idx];
2347 ///  reduce_function(GlobPtrs, reduce_data);
2348 static llvm::Value *emitListToGlobalReduceFunction(
2349     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2350     QualType ReductionArrayTy, SourceLocation Loc,
2351     const RecordDecl *TeamReductionRec,
2352     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2353         &VarFieldMap,
2354     llvm::Function *ReduceFn) {
2355   ASTContext &C = CGM.getContext();
2356 
2357   // Buffer: global reduction buffer.
2358   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2359                               C.VoidPtrTy, ImplicitParamDecl::Other);
2360   // Idx: index of the buffer.
2361   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2362                            ImplicitParamDecl::Other);
2363   // ReduceList: thread local Reduce list.
2364   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2365                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2366   FunctionArgList Args;
2367   Args.push_back(&BufferArg);
2368   Args.push_back(&IdxArg);
2369   Args.push_back(&ReduceListArg);
2370 
2371   const CGFunctionInfo &CGFI =
2372       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2373   auto *Fn = llvm::Function::Create(
2374       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2375       "_omp_reduction_list_to_global_reduce_func", &CGM.getModule());
2376   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2377   Fn->setDoesNotRecurse();
2378   CodeGenFunction CGF(CGM);
2379   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2380 
2381   CGBuilderTy &Bld = CGF.Builder;
2382 
2383   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2384   QualType StaticTy = C.getRecordType(TeamReductionRec);
2385   llvm::Type *LLVMReductionsBufferTy =
2386       CGM.getTypes().ConvertTypeForMem(StaticTy);
2387   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2388       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2389       LLVMReductionsBufferTy->getPointerTo());
2390 
2391   // 1. Build a list of reduction variables.
2392   // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
2393   Address ReductionList =
2394       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
2395   auto IPriv = Privates.begin();
2396   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2397                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2398                                               /*Volatile=*/false, C.IntTy,
2399                                               Loc)};
2400   unsigned Idx = 0;
2401   for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) {
2402     Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2403     // Global = Buffer.VD[Idx];
2404     const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl();
2405     const FieldDecl *FD = VarFieldMap.lookup(VD);
2406     LValue GlobLVal = CGF.EmitLValueForField(
2407         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2408     Address GlobAddr = GlobLVal.getAddress(CGF);
2409     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(
2410         GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs);
2411     CGF.EmitStoreOfScalar(BufferPtr, Elem, /*Volatile=*/false, C.VoidPtrTy);
2412     if ((*IPriv)->getType()->isVariablyModifiedType()) {
2413       // Store array size.
2414       ++Idx;
2415       Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2416       llvm::Value *Size = CGF.Builder.CreateIntCast(
2417           CGF.getVLASize(
2418                  CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
2419               .NumElts,
2420           CGF.SizeTy, /*isSigned=*/false);
2421       CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
2422                               Elem);
2423     }
2424   }
2425 
2426   // Call reduce_function(GlobalReduceList, ReduceList)
2427   llvm::Value *GlobalReduceList = ReductionList.getPointer();
2428   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2429   llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar(
2430       AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc);
2431   CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
2432       CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr});
2433   CGF.FinishFunction();
2434   return Fn;
2435 }
2436 
2437 /// This function emits a helper that copies all the reduction variables from
2438 /// the team into the provided global buffer for the reduction variables.
2439 ///
2440 /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
2441 ///   For all data entries D in reduce_data:
2442 ///     Copy buffer.D[Idx] to local D;
2443 static llvm::Value *emitGlobalToListCopyFunction(
2444     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2445     QualType ReductionArrayTy, SourceLocation Loc,
2446     const RecordDecl *TeamReductionRec,
2447     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2448         &VarFieldMap) {
2449   ASTContext &C = CGM.getContext();
2450 
2451   // Buffer: global reduction buffer.
2452   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2453                               C.VoidPtrTy, ImplicitParamDecl::Other);
2454   // Idx: index of the buffer.
2455   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2456                            ImplicitParamDecl::Other);
2457   // ReduceList: thread local Reduce list.
2458   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2459                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2460   FunctionArgList Args;
2461   Args.push_back(&BufferArg);
2462   Args.push_back(&IdxArg);
2463   Args.push_back(&ReduceListArg);
2464 
2465   const CGFunctionInfo &CGFI =
2466       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2467   auto *Fn = llvm::Function::Create(
2468       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2469       "_omp_reduction_global_to_list_copy_func", &CGM.getModule());
2470   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2471   Fn->setDoesNotRecurse();
2472   CodeGenFunction CGF(CGM);
2473   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2474 
2475   CGBuilderTy &Bld = CGF.Builder;
2476 
2477   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2478   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2479   llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
2480   Address LocalReduceList(
2481       Bld.CreatePointerBitCastOrAddrSpaceCast(
2482           CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
2483                                C.VoidPtrTy, Loc),
2484           ElemTy->getPointerTo()),
2485       ElemTy, CGF.getPointerAlign());
2486   QualType StaticTy = C.getRecordType(TeamReductionRec);
2487   llvm::Type *LLVMReductionsBufferTy =
2488       CGM.getTypes().ConvertTypeForMem(StaticTy);
2489   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2490       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2491       LLVMReductionsBufferTy->getPointerTo());
2492 
2493   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2494                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2495                                               /*Volatile=*/false, C.IntTy,
2496                                               Loc)};
2497   unsigned Idx = 0;
2498   for (const Expr *Private : Privates) {
2499     // Reduce element = LocalReduceList[i]
2500     Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
2501     llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
2502         ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
2503     // elemptr = ((CopyType*)(elemptrptr)) + I
2504     ElemTy = CGF.ConvertTypeForMem(Private->getType());
2505     ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2506         ElemPtrPtr, ElemTy->getPointerTo());
2507     Address ElemPtr =
2508         Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType()));
2509     const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl();
2510     // Global = Buffer.VD[Idx];
2511     const FieldDecl *FD = VarFieldMap.lookup(VD);
2512     LValue GlobLVal = CGF.EmitLValueForField(
2513         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2514     Address GlobAddr = GlobLVal.getAddress(CGF);
2515     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(),
2516                                                    GlobAddr.getPointer(), Idxs);
2517     GlobLVal.setAddress(Address(BufferPtr,
2518                                 CGF.ConvertTypeForMem(Private->getType()),
2519                                 GlobAddr.getAlignment()));
2520     switch (CGF.getEvaluationKind(Private->getType())) {
2521     case TEK_Scalar: {
2522       llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc);
2523       CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType(),
2524                             LValueBaseInfo(AlignmentSource::Type),
2525                             TBAAAccessInfo());
2526       break;
2527     }
2528     case TEK_Complex: {
2529       CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc);
2530       CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()),
2531                              /*isInit=*/false);
2532       break;
2533     }
2534     case TEK_Aggregate:
2535       CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()),
2536                             GlobLVal, Private->getType(),
2537                             AggValueSlot::DoesNotOverlap);
2538       break;
2539     }
2540     ++Idx;
2541   }
2542 
2543   CGF.FinishFunction();
2544   return Fn;
2545 }
2546 
2547 /// This function emits a helper that reduces all the reduction variables from
2548 /// the team into the provided global buffer for the reduction variables.
2549 ///
2550 /// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data)
2551 ///  void *GlobPtrs[];
2552 ///  GlobPtrs[0] = (void*)&buffer.D0[Idx];
2553 ///  ...
2554 ///  GlobPtrs[N] = (void*)&buffer.DN[Idx];
2555 ///  reduce_function(reduce_data, GlobPtrs);
2556 static llvm::Value *emitGlobalToListReduceFunction(
2557     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2558     QualType ReductionArrayTy, SourceLocation Loc,
2559     const RecordDecl *TeamReductionRec,
2560     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2561         &VarFieldMap,
2562     llvm::Function *ReduceFn) {
2563   ASTContext &C = CGM.getContext();
2564 
2565   // Buffer: global reduction buffer.
2566   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2567                               C.VoidPtrTy, ImplicitParamDecl::Other);
2568   // Idx: index of the buffer.
2569   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2570                            ImplicitParamDecl::Other);
2571   // ReduceList: thread local Reduce list.
2572   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2573                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2574   FunctionArgList Args;
2575   Args.push_back(&BufferArg);
2576   Args.push_back(&IdxArg);
2577   Args.push_back(&ReduceListArg);
2578 
2579   const CGFunctionInfo &CGFI =
2580       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2581   auto *Fn = llvm::Function::Create(
2582       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2583       "_omp_reduction_global_to_list_reduce_func", &CGM.getModule());
2584   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2585   Fn->setDoesNotRecurse();
2586   CodeGenFunction CGF(CGM);
2587   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2588 
2589   CGBuilderTy &Bld = CGF.Builder;
2590 
2591   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2592   QualType StaticTy = C.getRecordType(TeamReductionRec);
2593   llvm::Type *LLVMReductionsBufferTy =
2594       CGM.getTypes().ConvertTypeForMem(StaticTy);
2595   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2596       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2597       LLVMReductionsBufferTy->getPointerTo());
2598 
2599   // 1. Build a list of reduction variables.
2600   // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
2601   Address ReductionList =
2602       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
2603   auto IPriv = Privates.begin();
2604   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2605                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2606                                               /*Volatile=*/false, C.IntTy,
2607                                               Loc)};
2608   unsigned Idx = 0;
2609   for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) {
2610     Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2611     // Global = Buffer.VD[Idx];
2612     const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl();
2613     const FieldDecl *FD = VarFieldMap.lookup(VD);
2614     LValue GlobLVal = CGF.EmitLValueForField(
2615         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2616     Address GlobAddr = GlobLVal.getAddress(CGF);
2617     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(
2618         GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs);
2619     CGF.EmitStoreOfScalar(BufferPtr, Elem, /*Volatile=*/false, C.VoidPtrTy);
2620     if ((*IPriv)->getType()->isVariablyModifiedType()) {
2621       // Store array size.
2622       ++Idx;
2623       Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2624       llvm::Value *Size = CGF.Builder.CreateIntCast(
2625           CGF.getVLASize(
2626                  CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
2627               .NumElts,
2628           CGF.SizeTy, /*isSigned=*/false);
2629       CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
2630                               Elem);
2631     }
2632   }
2633 
2634   // Call reduce_function(ReduceList, GlobalReduceList)
2635   llvm::Value *GlobalReduceList = ReductionList.getPointer();
2636   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2637   llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar(
2638       AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc);
2639   CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
2640       CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList});
2641   CGF.FinishFunction();
2642   return Fn;
2643 }
2644 
2645 ///
2646 /// Design of OpenMP reductions on the GPU
2647 ///
2648 /// Consider a typical OpenMP program with one or more reduction
2649 /// clauses:
2650 ///
2651 /// float foo;
2652 /// double bar;
2653 /// #pragma omp target teams distribute parallel for \
2654 ///             reduction(+:foo) reduction(*:bar)
2655 /// for (int i = 0; i < N; i++) {
2656 ///   foo += A[i]; bar *= B[i];
2657 /// }
2658 ///
2659 /// where 'foo' and 'bar' are reduced across all OpenMP threads in
2660 /// all teams.  In our OpenMP implementation on the NVPTX device an
2661 /// OpenMP team is mapped to a CUDA threadblock and OpenMP threads
2662 /// within a team are mapped to CUDA threads within a threadblock.
2663 /// Our goal is to efficiently aggregate values across all OpenMP
2664 /// threads such that:
2665 ///
2666 ///   - the compiler and runtime are logically concise, and
2667 ///   - the reduction is performed efficiently in a hierarchical
2668 ///     manner as follows: within OpenMP threads in the same warp,
2669 ///     across warps in a threadblock, and finally across teams on
2670 ///     the NVPTX device.
2671 ///
2672 /// Introduction to Decoupling
2673 ///
2674 /// We would like to decouple the compiler and the runtime so that the
2675 /// latter is ignorant of the reduction variables (number, data types)
2676 /// and the reduction operators.  This allows a simpler interface
2677 /// and implementation while still attaining good performance.
2678 ///
2679 /// Pseudocode for the aforementioned OpenMP program generated by the
2680 /// compiler is as follows:
2681 ///
2682 /// 1. Create private copies of reduction variables on each OpenMP
2683 ///    thread: 'foo_private', 'bar_private'
2684 /// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned
2685 ///    to it and writes the result in 'foo_private' and 'bar_private'
2686 ///    respectively.
2687 /// 3. Call the OpenMP runtime on the GPU to reduce within a team
2688 ///    and store the result on the team master:
2689 ///
2690 ///     __kmpc_nvptx_parallel_reduce_nowait_v2(...,
2691 ///        reduceData, shuffleReduceFn, interWarpCpyFn)
2692 ///
2693 ///     where:
2694 ///       struct ReduceData {
2695 ///         double *foo;
2696 ///         double *bar;
2697 ///       } reduceData
2698 ///       reduceData.foo = &foo_private
2699 ///       reduceData.bar = &bar_private
2700 ///
2701 ///     'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two
2702 ///     auxiliary functions generated by the compiler that operate on
2703 ///     variables of type 'ReduceData'.  They aid the runtime perform
2704 ///     algorithmic steps in a data agnostic manner.
2705 ///
2706 ///     'shuffleReduceFn' is a pointer to a function that reduces data
2707 ///     of type 'ReduceData' across two OpenMP threads (lanes) in the
2708 ///     same warp.  It takes the following arguments as input:
2709 ///
2710 ///     a. variable of type 'ReduceData' on the calling lane,
2711 ///     b. its lane_id,
2712 ///     c. an offset relative to the current lane_id to generate a
2713 ///        remote_lane_id.  The remote lane contains the second
2714 ///        variable of type 'ReduceData' that is to be reduced.
2715 ///     d. an algorithm version parameter determining which reduction
2716 ///        algorithm to use.
2717 ///
2718 ///     'shuffleReduceFn' retrieves data from the remote lane using
2719 ///     efficient GPU shuffle intrinsics and reduces, using the
2720 ///     algorithm specified by the 4th parameter, the two operands
2721 ///     element-wise.  The result is written to the first operand.
2722 ///
2723 ///     Different reduction algorithms are implemented in different
2724 ///     runtime functions, all calling 'shuffleReduceFn' to perform
2725 ///     the essential reduction step.  Therefore, based on the 4th
2726 ///     parameter, this function behaves slightly differently to
2727 ///     cooperate with the runtime to ensure correctness under
2728 ///     different circumstances.
2729 ///
2730 ///     'InterWarpCpyFn' is a pointer to a function that transfers
2731 ///     reduced variables across warps.  It tunnels, through CUDA
2732 ///     shared memory, the thread-private data of type 'ReduceData'
2733 ///     from lane 0 of each warp to a lane in the first warp.
2734 /// 4. Call the OpenMP runtime on the GPU to reduce across teams.
2735 ///    The last team writes the global reduced value to memory.
2736 ///
2737 ///     ret = __kmpc_nvptx_teams_reduce_nowait(...,
2738 ///             reduceData, shuffleReduceFn, interWarpCpyFn,
2739 ///             scratchpadCopyFn, loadAndReduceFn)
2740 ///
2741 ///     'scratchpadCopyFn' is a helper that stores reduced
2742 ///     data from the team master to a scratchpad array in
2743 ///     global memory.
2744 ///
2745 ///     'loadAndReduceFn' is a helper that loads data from
2746 ///     the scratchpad array and reduces it with the input
2747 ///     operand.
2748 ///
2749 ///     These compiler generated functions hide address
2750 ///     calculation and alignment information from the runtime.
2751 /// 5. if ret == 1:
2752 ///     The team master of the last team stores the reduced
2753 ///     result to the globals in memory.
2754 ///     foo += reduceData.foo; bar *= reduceData.bar
2755 ///
2756 ///
2757 /// Warp Reduction Algorithms
2758 ///
2759 /// On the warp level, we have three algorithms implemented in the
2760 /// OpenMP runtime depending on the number of active lanes:
2761 ///
2762 /// Full Warp Reduction
2763 ///
2764 /// The reduce algorithm within a warp where all lanes are active
2765 /// is implemented in the runtime as follows:
2766 ///
2767 /// full_warp_reduce(void *reduce_data,
2768 ///                  kmp_ShuffleReductFctPtr ShuffleReduceFn) {
2769 ///   for (int offset = WARPSIZE/2; offset > 0; offset /= 2)
2770 ///     ShuffleReduceFn(reduce_data, 0, offset, 0);
2771 /// }
2772 ///
2773 /// The algorithm completes in log(2, WARPSIZE) steps.
2774 ///
2775 /// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is
2776 /// not used therefore we save instructions by not retrieving lane_id
2777 /// from the corresponding special registers.  The 4th parameter, which
2778 /// represents the version of the algorithm being used, is set to 0 to
2779 /// signify full warp reduction.
2780 ///
2781 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
2782 ///
2783 /// #reduce_elem refers to an element in the local lane's data structure
2784 /// #remote_elem is retrieved from a remote lane
2785 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
2786 /// reduce_elem = reduce_elem REDUCE_OP remote_elem;
2787 ///
2788 /// Contiguous Partial Warp Reduction
2789 ///
2790 /// This reduce algorithm is used within a warp where only the first
2791 /// 'n' (n <= WARPSIZE) lanes are active.  It is typically used when the
2792 /// number of OpenMP threads in a parallel region is not a multiple of
2793 /// WARPSIZE.  The algorithm is implemented in the runtime as follows:
2794 ///
2795 /// void
2796 /// contiguous_partial_reduce(void *reduce_data,
2797 ///                           kmp_ShuffleReductFctPtr ShuffleReduceFn,
2798 ///                           int size, int lane_id) {
2799 ///   int curr_size;
2800 ///   int offset;
2801 ///   curr_size = size;
2802 ///   mask = curr_size/2;
2803 ///   while (offset>0) {
2804 ///     ShuffleReduceFn(reduce_data, lane_id, offset, 1);
2805 ///     curr_size = (curr_size+1)/2;
2806 ///     offset = curr_size/2;
2807 ///   }
2808 /// }
2809 ///
2810 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
2811 ///
2812 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
2813 /// if (lane_id < offset)
2814 ///     reduce_elem = reduce_elem REDUCE_OP remote_elem
2815 /// else
2816 ///     reduce_elem = remote_elem
2817 ///
2818 /// This algorithm assumes that the data to be reduced are located in a
2819 /// contiguous subset of lanes starting from the first.  When there is
2820 /// an odd number of active lanes, the data in the last lane is not
2821 /// aggregated with any other lane's dat but is instead copied over.
2822 ///
2823 /// Dispersed Partial Warp Reduction
2824 ///
2825 /// This algorithm is used within a warp when any discontiguous subset of
2826 /// lanes are active.  It is used to implement the reduction operation
2827 /// across lanes in an OpenMP simd region or in a nested parallel region.
2828 ///
2829 /// void
2830 /// dispersed_partial_reduce(void *reduce_data,
2831 ///                          kmp_ShuffleReductFctPtr ShuffleReduceFn) {
2832 ///   int size, remote_id;
2833 ///   int logical_lane_id = number_of_active_lanes_before_me() * 2;
2834 ///   do {
2835 ///       remote_id = next_active_lane_id_right_after_me();
2836 ///       # the above function returns 0 of no active lane
2837 ///       # is present right after the current lane.
2838 ///       size = number_of_active_lanes_in_this_warp();
2839 ///       logical_lane_id /= 2;
2840 ///       ShuffleReduceFn(reduce_data, logical_lane_id,
2841 ///                       remote_id-1-threadIdx.x, 2);
2842 ///   } while (logical_lane_id % 2 == 0 && size > 1);
2843 /// }
2844 ///
2845 /// There is no assumption made about the initial state of the reduction.
2846 /// Any number of lanes (>=1) could be active at any position.  The reduction
2847 /// result is returned in the first active lane.
2848 ///
2849 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
2850 ///
2851 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
2852 /// if (lane_id % 2 == 0 && offset > 0)
2853 ///     reduce_elem = reduce_elem REDUCE_OP remote_elem
2854 /// else
2855 ///     reduce_elem = remote_elem
2856 ///
2857 ///
2858 /// Intra-Team Reduction
2859 ///
2860 /// This function, as implemented in the runtime call
2861 /// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP
2862 /// threads in a team.  It first reduces within a warp using the
2863 /// aforementioned algorithms.  We then proceed to gather all such
2864 /// reduced values at the first warp.
2865 ///
2866 /// The runtime makes use of the function 'InterWarpCpyFn', which copies
2867 /// data from each of the "warp master" (zeroth lane of each warp, where
2868 /// warp-reduced data is held) to the zeroth warp.  This step reduces (in
2869 /// a mathematical sense) the problem of reduction across warp masters in
2870 /// a block to the problem of warp reduction.
2871 ///
2872 ///
2873 /// Inter-Team Reduction
2874 ///
2875 /// Once a team has reduced its data to a single value, it is stored in
2876 /// a global scratchpad array.  Since each team has a distinct slot, this
2877 /// can be done without locking.
2878 ///
2879 /// The last team to write to the scratchpad array proceeds to reduce the
2880 /// scratchpad array.  One or more workers in the last team use the helper
2881 /// 'loadAndReduceDataFn' to load and reduce values from the array, i.e.,
2882 /// the k'th worker reduces every k'th element.
2883 ///
2884 /// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to
2885 /// reduce across workers and compute a globally reduced value.
2886 ///
2887 void CGOpenMPRuntimeGPU::emitReduction(
2888     CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates,
2889     ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs,
2890     ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) {
2891   if (!CGF.HaveInsertPoint())
2892     return;
2893 
2894   bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind);
2895 #ifndef NDEBUG
2896   bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind);
2897 #endif
2898 
2899   if (Options.SimpleReduction) {
2900     assert(!TeamsReduction && !ParallelReduction &&
2901            "Invalid reduction selection in emitReduction.");
2902     CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs,
2903                                    ReductionOps, Options);
2904     return;
2905   }
2906 
2907   assert((TeamsReduction || ParallelReduction) &&
2908          "Invalid reduction selection in emitReduction.");
2909 
2910   // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList),
2911   // RedList, shuffle_reduce_func, interwarp_copy_func);
2912   // or
2913   // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>);
2914   llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
2915   llvm::Value *ThreadId = getThreadID(CGF, Loc);
2916 
2917   llvm::Value *Res;
2918   ASTContext &C = CGM.getContext();
2919   // 1. Build a list of reduction variables.
2920   // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
2921   auto Size = RHSExprs.size();
2922   for (const Expr *E : Privates) {
2923     if (E->getType()->isVariablyModifiedType())
2924       // Reserve place for array size.
2925       ++Size;
2926   }
2927   llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size);
2928   QualType ReductionArrayTy =
2929       C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal,
2930                              /*IndexTypeQuals=*/0);
2931   Address ReductionList =
2932       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
2933   auto IPriv = Privates.begin();
2934   unsigned Idx = 0;
2935   for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) {
2936     Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2937     CGF.Builder.CreateStore(
2938         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2939             CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy),
2940         Elem);
2941     if ((*IPriv)->getType()->isVariablyModifiedType()) {
2942       // Store array size.
2943       ++Idx;
2944       Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2945       llvm::Value *Size = CGF.Builder.CreateIntCast(
2946           CGF.getVLASize(
2947                  CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
2948               .NumElts,
2949           CGF.SizeTy, /*isSigned=*/false);
2950       CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
2951                               Elem);
2952     }
2953   }
2954 
2955   llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2956       ReductionList.getPointer(), CGF.VoidPtrTy);
2957   llvm::Function *ReductionFn = emitReductionFunction(
2958       CGF.CurFn->getName(), Loc, CGF.ConvertTypeForMem(ReductionArrayTy),
2959       Privates, LHSExprs, RHSExprs, ReductionOps);
2960   llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy);
2961   llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction(
2962       CGM, Privates, ReductionArrayTy, ReductionFn, Loc);
2963   llvm::Value *InterWarpCopyFn =
2964       emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc);
2965 
2966   if (ParallelReduction) {
2967     llvm::Value *Args[] = {RTLoc,
2968                            ThreadId,
2969                            CGF.Builder.getInt32(RHSExprs.size()),
2970                            ReductionArrayTySize,
2971                            RL,
2972                            ShuffleAndReduceFn,
2973                            InterWarpCopyFn};
2974 
2975     Res = CGF.EmitRuntimeCall(
2976         OMPBuilder.getOrCreateRuntimeFunction(
2977             CGM.getModule(), OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2),
2978         Args);
2979   } else {
2980     assert(TeamsReduction && "expected teams reduction.");
2981     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap;
2982     llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size());
2983     int Cnt = 0;
2984     for (const Expr *DRE : Privates) {
2985       PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl();
2986       ++Cnt;
2987     }
2988     const RecordDecl *TeamReductionRec = ::buildRecordForGlobalizedVars(
2989         CGM.getContext(), PrivatesReductions, std::nullopt, VarFieldMap,
2990         C.getLangOpts().OpenMPCUDAReductionBufNum);
2991     TeamsReductions.push_back(TeamReductionRec);
2992     if (!KernelTeamsReductionPtr) {
2993       KernelTeamsReductionPtr = new llvm::GlobalVariable(
2994           CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/true,
2995           llvm::GlobalValue::InternalLinkage, nullptr,
2996           "_openmp_teams_reductions_buffer_$_$ptr");
2997     }
2998     llvm::Value *GlobalBufferPtr = CGF.EmitLoadOfScalar(
2999         Address(KernelTeamsReductionPtr, CGF.VoidPtrTy, CGM.getPointerAlign()),
3000         /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc);
3001     llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction(
3002         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap);
3003     llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction(
3004         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap,
3005         ReductionFn);
3006     llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction(
3007         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap);
3008     llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction(
3009         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap,
3010         ReductionFn);
3011 
3012     llvm::Value *Args[] = {
3013         RTLoc,
3014         ThreadId,
3015         GlobalBufferPtr,
3016         CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum),
3017         RL,
3018         ShuffleAndReduceFn,
3019         InterWarpCopyFn,
3020         GlobalToBufferCpyFn,
3021         GlobalToBufferRedFn,
3022         BufferToGlobalCpyFn,
3023         BufferToGlobalRedFn};
3024 
3025     Res = CGF.EmitRuntimeCall(
3026         OMPBuilder.getOrCreateRuntimeFunction(
3027             CGM.getModule(), OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2),
3028         Args);
3029   }
3030 
3031   // 5. Build if (res == 1)
3032   llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done");
3033   llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then");
3034   llvm::Value *Cond = CGF.Builder.CreateICmpEQ(
3035       Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1));
3036   CGF.Builder.CreateCondBr(Cond, ThenBB, ExitBB);
3037 
3038   // 6. Build then branch: where we have reduced values in the master
3039   //    thread in each team.
3040   //    __kmpc_end_reduce{_nowait}(<gtid>);
3041   //    break;
3042   CGF.EmitBlock(ThenBB);
3043 
3044   // Add emission of __kmpc_end_reduce{_nowait}(<gtid>);
3045   auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps,
3046                     this](CodeGenFunction &CGF, PrePostActionTy &Action) {
3047     auto IPriv = Privates.begin();
3048     auto ILHS = LHSExprs.begin();
3049     auto IRHS = RHSExprs.begin();
3050     for (const Expr *E : ReductionOps) {
3051       emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
3052                                   cast<DeclRefExpr>(*IRHS));
3053       ++IPriv;
3054       ++ILHS;
3055       ++IRHS;
3056     }
3057   };
3058   llvm::Value *EndArgs[] = {ThreadId};
3059   RegionCodeGenTy RCG(CodeGen);
3060   NVPTXActionTy Action(
3061       nullptr, std::nullopt,
3062       OMPBuilder.getOrCreateRuntimeFunction(
3063           CGM.getModule(), OMPRTL___kmpc_nvptx_end_reduce_nowait),
3064       EndArgs);
3065   RCG.setAction(Action);
3066   RCG(CGF);
3067   // There is no need to emit line number for unconditional branch.
3068   (void)ApplyDebugLocation::CreateEmpty(CGF);
3069   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
3070 }
3071 
3072 const VarDecl *
3073 CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD,
3074                                        const VarDecl *NativeParam) const {
3075   if (!NativeParam->getType()->isReferenceType())
3076     return NativeParam;
3077   QualType ArgType = NativeParam->getType();
3078   QualifierCollector QC;
3079   const Type *NonQualTy = QC.strip(ArgType);
3080   QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
3081   if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) {
3082     if (Attr->getCaptureKind() == OMPC_map) {
3083       PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy,
3084                                                         LangAS::opencl_global);
3085     }
3086   }
3087   ArgType = CGM.getContext().getPointerType(PointeeTy);
3088   QC.addRestrict();
3089   enum { NVPTX_local_addr = 5 };
3090   QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr));
3091   ArgType = QC.apply(CGM.getContext(), ArgType);
3092   if (isa<ImplicitParamDecl>(NativeParam))
3093     return ImplicitParamDecl::Create(
3094         CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(),
3095         NativeParam->getIdentifier(), ArgType, ImplicitParamDecl::Other);
3096   return ParmVarDecl::Create(
3097       CGM.getContext(),
3098       const_cast<DeclContext *>(NativeParam->getDeclContext()),
3099       NativeParam->getBeginLoc(), NativeParam->getLocation(),
3100       NativeParam->getIdentifier(), ArgType,
3101       /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr);
3102 }
3103 
3104 Address
3105 CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF,
3106                                           const VarDecl *NativeParam,
3107                                           const VarDecl *TargetParam) const {
3108   assert(NativeParam != TargetParam &&
3109          NativeParam->getType()->isReferenceType() &&
3110          "Native arg must not be the same as target arg.");
3111   Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam);
3112   QualType NativeParamType = NativeParam->getType();
3113   QualifierCollector QC;
3114   const Type *NonQualTy = QC.strip(NativeParamType);
3115   QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
3116   unsigned NativePointeeAddrSpace =
3117       CGF.getTypes().getTargetAddressSpace(NativePointeeTy);
3118   QualType TargetTy = TargetParam->getType();
3119   llvm::Value *TargetAddr = CGF.EmitLoadOfScalar(LocalAddr, /*Volatile=*/false,
3120                                                  TargetTy, SourceLocation());
3121   // First cast to generic.
3122   TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3123       TargetAddr,
3124       llvm::PointerType::get(CGF.getLLVMContext(), /*AddrSpace=*/0));
3125   // Cast from generic to native address space.
3126   TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3127       TargetAddr,
3128       llvm::PointerType::get(CGF.getLLVMContext(), NativePointeeAddrSpace));
3129   Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType);
3130   CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false,
3131                         NativeParamType);
3132   return NativeParamAddr;
3133 }
3134 
3135 void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall(
3136     CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn,
3137     ArrayRef<llvm::Value *> Args) const {
3138   SmallVector<llvm::Value *, 4> TargetArgs;
3139   TargetArgs.reserve(Args.size());
3140   auto *FnType = OutlinedFn.getFunctionType();
3141   for (unsigned I = 0, E = Args.size(); I < E; ++I) {
3142     if (FnType->isVarArg() && FnType->getNumParams() <= I) {
3143       TargetArgs.append(std::next(Args.begin(), I), Args.end());
3144       break;
3145     }
3146     llvm::Type *TargetType = FnType->getParamType(I);
3147     llvm::Value *NativeArg = Args[I];
3148     if (!TargetType->isPointerTy()) {
3149       TargetArgs.emplace_back(NativeArg);
3150       continue;
3151     }
3152     llvm::Value *TargetArg = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3153         NativeArg,
3154         llvm::PointerType::get(CGF.getLLVMContext(), /*AddrSpace*/ 0));
3155     TargetArgs.emplace_back(
3156         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType));
3157   }
3158   CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs);
3159 }
3160 
3161 /// Emit function which wraps the outline parallel region
3162 /// and controls the arguments which are passed to this function.
3163 /// The wrapper ensures that the outlined function is called
3164 /// with the correct arguments when data is shared.
3165 llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper(
3166     llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) {
3167   ASTContext &Ctx = CGM.getContext();
3168   const auto &CS = *D.getCapturedStmt(OMPD_parallel);
3169 
3170   // Create a function that takes as argument the source thread.
3171   FunctionArgList WrapperArgs;
3172   QualType Int16QTy =
3173       Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false);
3174   QualType Int32QTy =
3175       Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false);
3176   ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
3177                                      /*Id=*/nullptr, Int16QTy,
3178                                      ImplicitParamDecl::Other);
3179   ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
3180                                /*Id=*/nullptr, Int32QTy,
3181                                ImplicitParamDecl::Other);
3182   WrapperArgs.emplace_back(&ParallelLevelArg);
3183   WrapperArgs.emplace_back(&WrapperArg);
3184 
3185   const CGFunctionInfo &CGFI =
3186       CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs);
3187 
3188   auto *Fn = llvm::Function::Create(
3189       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
3190       Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule());
3191 
3192   // Ensure we do not inline the function. This is trivially true for the ones
3193   // passed to __kmpc_fork_call but the ones calles in serialized regions
3194   // could be inlined. This is not a perfect but it is closer to the invariant
3195   // we want, namely, every data environment starts with a new function.
3196   // TODO: We should pass the if condition to the runtime function and do the
3197   //       handling there. Much cleaner code.
3198   Fn->addFnAttr(llvm::Attribute::NoInline);
3199 
3200   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
3201   Fn->setLinkage(llvm::GlobalValue::InternalLinkage);
3202   Fn->setDoesNotRecurse();
3203 
3204   CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
3205   CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs,
3206                     D.getBeginLoc(), D.getBeginLoc());
3207 
3208   const auto *RD = CS.getCapturedRecordDecl();
3209   auto CurField = RD->field_begin();
3210 
3211   Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
3212                                                       /*Name=*/".zero.addr");
3213   CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr);
3214   // Get the array of arguments.
3215   SmallVector<llvm::Value *, 8> Args;
3216 
3217   Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer());
3218   Args.emplace_back(ZeroAddr.getPointer());
3219 
3220   CGBuilderTy &Bld = CGF.Builder;
3221   auto CI = CS.capture_begin();
3222 
3223   // Use global memory for data sharing.
3224   // Handle passing of global args to workers.
3225   Address GlobalArgs =
3226       CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args");
3227   llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer();
3228   llvm::Value *DataSharingArgs[] = {GlobalArgsPtr};
3229   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
3230                           CGM.getModule(), OMPRTL___kmpc_get_shared_variables),
3231                       DataSharingArgs);
3232 
3233   // Retrieve the shared variables from the list of references returned
3234   // by the runtime. Pass the variables to the outlined function.
3235   Address SharedArgListAddress = Address::invalid();
3236   if (CS.capture_size() > 0 ||
3237       isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
3238     SharedArgListAddress = CGF.EmitLoadOfPointer(
3239         GlobalArgs, CGF.getContext()
3240                         .getPointerType(CGF.getContext().VoidPtrTy)
3241                         .castAs<PointerType>());
3242   }
3243   unsigned Idx = 0;
3244   if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
3245     Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
3246     Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
3247         Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy);
3248     llvm::Value *LB = CGF.EmitLoadOfScalar(
3249         TypedAddress,
3250         /*Volatile=*/false,
3251         CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
3252         cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc());
3253     Args.emplace_back(LB);
3254     ++Idx;
3255     Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
3256     TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
3257         Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy);
3258     llvm::Value *UB = CGF.EmitLoadOfScalar(
3259         TypedAddress,
3260         /*Volatile=*/false,
3261         CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
3262         cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc());
3263     Args.emplace_back(UB);
3264     ++Idx;
3265   }
3266   if (CS.capture_size() > 0) {
3267     ASTContext &CGFContext = CGF.getContext();
3268     for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) {
3269       QualType ElemTy = CurField->getType();
3270       Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx);
3271       Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
3272           Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)),
3273           CGF.ConvertTypeForMem(ElemTy));
3274       llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress,
3275                                               /*Volatile=*/false,
3276                                               CGFContext.getPointerType(ElemTy),
3277                                               CI->getLocation());
3278       if (CI->capturesVariableByCopy() &&
3279           !CI->getCapturedVar()->getType()->isAnyPointerType()) {
3280         Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(),
3281                               CI->getLocation());
3282       }
3283       Args.emplace_back(Arg);
3284     }
3285   }
3286 
3287   emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args);
3288   CGF.FinishFunction();
3289   return Fn;
3290 }
3291 
3292 void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF,
3293                                               const Decl *D) {
3294   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic)
3295     return;
3296 
3297   assert(D && "Expected function or captured|block decl.");
3298   assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 &&
3299          "Function is registered already.");
3300   assert((!TeamAndReductions.first || TeamAndReductions.first == D) &&
3301          "Team is set but not processed.");
3302   const Stmt *Body = nullptr;
3303   bool NeedToDelayGlobalization = false;
3304   if (const auto *FD = dyn_cast<FunctionDecl>(D)) {
3305     Body = FD->getBody();
3306   } else if (const auto *BD = dyn_cast<BlockDecl>(D)) {
3307     Body = BD->getBody();
3308   } else if (const auto *CD = dyn_cast<CapturedDecl>(D)) {
3309     Body = CD->getBody();
3310     NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP;
3311     if (NeedToDelayGlobalization &&
3312         getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
3313       return;
3314   }
3315   if (!Body)
3316     return;
3317   CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second);
3318   VarChecker.Visit(Body);
3319   const RecordDecl *GlobalizedVarsRecord =
3320       VarChecker.getGlobalizedRecord(IsInTTDRegion);
3321   TeamAndReductions.first = nullptr;
3322   TeamAndReductions.second.clear();
3323   ArrayRef<const ValueDecl *> EscapedVariableLengthDecls =
3324       VarChecker.getEscapedVariableLengthDecls();
3325   ArrayRef<const ValueDecl *> DelayedVariableLengthDecls =
3326       VarChecker.getDelayedVariableLengthDecls();
3327   if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty() &&
3328       DelayedVariableLengthDecls.empty())
3329     return;
3330   auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
3331   I->getSecond().MappedParams =
3332       std::make_unique<CodeGenFunction::OMPMapVars>();
3333   I->getSecond().EscapedParameters.insert(
3334       VarChecker.getEscapedParameters().begin(),
3335       VarChecker.getEscapedParameters().end());
3336   I->getSecond().EscapedVariableLengthDecls.append(
3337       EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end());
3338   I->getSecond().DelayedVariableLengthDecls.append(
3339       DelayedVariableLengthDecls.begin(), DelayedVariableLengthDecls.end());
3340   DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
3341   for (const ValueDecl *VD : VarChecker.getEscapedDecls()) {
3342     assert(VD->isCanonicalDecl() && "Expected canonical declaration");
3343     Data.insert(std::make_pair(VD, MappedVarData()));
3344   }
3345   if (!NeedToDelayGlobalization) {
3346     emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true);
3347     struct GlobalizationScope final : EHScopeStack::Cleanup {
3348       GlobalizationScope() = default;
3349 
3350       void Emit(CodeGenFunction &CGF, Flags flags) override {
3351         static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
3352             .emitGenericVarsEpilog(CGF, /*WithSPMDCheck=*/true);
3353       }
3354     };
3355     CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup);
3356   }
3357 }
3358 
3359 Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF,
3360                                                         const VarDecl *VD) {
3361   if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) {
3362     const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
3363     auto AS = LangAS::Default;
3364     switch (A->getAllocatorType()) {
3365       // Use the default allocator here as by default local vars are
3366       // threadlocal.
3367     case OMPAllocateDeclAttr::OMPNullMemAlloc:
3368     case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
3369     case OMPAllocateDeclAttr::OMPThreadMemAlloc:
3370     case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
3371     case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
3372       // Follow the user decision - use default allocation.
3373       return Address::invalid();
3374     case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
3375       // TODO: implement aupport for user-defined allocators.
3376       return Address::invalid();
3377     case OMPAllocateDeclAttr::OMPConstMemAlloc:
3378       AS = LangAS::cuda_constant;
3379       break;
3380     case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
3381       AS = LangAS::cuda_shared;
3382       break;
3383     case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
3384     case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
3385       break;
3386     }
3387     llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
3388     auto *GV = new llvm::GlobalVariable(
3389         CGM.getModule(), VarTy, /*isConstant=*/false,
3390         llvm::GlobalValue::InternalLinkage, llvm::PoisonValue::get(VarTy),
3391         VD->getName(),
3392         /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
3393         CGM.getContext().getTargetAddressSpace(AS));
3394     CharUnits Align = CGM.getContext().getDeclAlign(VD);
3395     GV->setAlignment(Align.getAsAlign());
3396     return Address(
3397         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3398             GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace(
3399                     VD->getType().getAddressSpace()))),
3400         VarTy, Align);
3401   }
3402 
3403   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic)
3404     return Address::invalid();
3405 
3406   VD = VD->getCanonicalDecl();
3407   auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
3408   if (I == FunctionGlobalizedDecls.end())
3409     return Address::invalid();
3410   auto VDI = I->getSecond().LocalVarData.find(VD);
3411   if (VDI != I->getSecond().LocalVarData.end())
3412     return VDI->second.PrivateAddr;
3413   if (VD->hasAttrs()) {
3414     for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()),
3415          E(VD->attr_end());
3416          IT != E; ++IT) {
3417       auto VDI = I->getSecond().LocalVarData.find(
3418           cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl())
3419               ->getCanonicalDecl());
3420       if (VDI != I->getSecond().LocalVarData.end())
3421         return VDI->second.PrivateAddr;
3422     }
3423   }
3424 
3425   return Address::invalid();
3426 }
3427 
3428 void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) {
3429   FunctionGlobalizedDecls.erase(CGF.CurFn);
3430   CGOpenMPRuntime::functionFinished(CGF);
3431 }
3432 
3433 void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk(
3434     CodeGenFunction &CGF, const OMPLoopDirective &S,
3435     OpenMPDistScheduleClauseKind &ScheduleKind,
3436     llvm::Value *&Chunk) const {
3437   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
3438   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
3439     ScheduleKind = OMPC_DIST_SCHEDULE_static;
3440     Chunk = CGF.EmitScalarConversion(
3441         RT.getGPUNumThreads(CGF),
3442         CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
3443         S.getIterationVariable()->getType(), S.getBeginLoc());
3444     return;
3445   }
3446   CGOpenMPRuntime::getDefaultDistScheduleAndChunk(
3447       CGF, S, ScheduleKind, Chunk);
3448 }
3449 
3450 void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk(
3451     CodeGenFunction &CGF, const OMPLoopDirective &S,
3452     OpenMPScheduleClauseKind &ScheduleKind,
3453     const Expr *&ChunkExpr) const {
3454   ScheduleKind = OMPC_SCHEDULE_static;
3455   // Chunk size is 1 in this case.
3456   llvm::APInt ChunkSize(32, 1);
3457   ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize,
3458       CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
3459       SourceLocation());
3460 }
3461 
3462 void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas(
3463     CodeGenFunction &CGF, const OMPExecutableDirective &D) const {
3464   assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&
3465          " Expected target-based directive.");
3466   const CapturedStmt *CS = D.getCapturedStmt(OMPD_target);
3467   for (const CapturedStmt::Capture &C : CS->captures()) {
3468     // Capture variables captured by reference in lambdas for target-based
3469     // directives.
3470     if (!C.capturesVariable())
3471       continue;
3472     const VarDecl *VD = C.getCapturedVar();
3473     const auto *RD = VD->getType()
3474                          .getCanonicalType()
3475                          .getNonReferenceType()
3476                          ->getAsCXXRecordDecl();
3477     if (!RD || !RD->isLambda())
3478       continue;
3479     Address VDAddr = CGF.GetAddrOfLocalVar(VD);
3480     LValue VDLVal;
3481     if (VD->getType().getCanonicalType()->isReferenceType())
3482       VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType());
3483     else
3484       VDLVal = CGF.MakeAddrLValue(
3485           VDAddr, VD->getType().getCanonicalType().getNonReferenceType());
3486     llvm::DenseMap<const ValueDecl *, FieldDecl *> Captures;
3487     FieldDecl *ThisCapture = nullptr;
3488     RD->getCaptureFields(Captures, ThisCapture);
3489     if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) {
3490       LValue ThisLVal =
3491           CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture);
3492       llvm::Value *CXXThis = CGF.LoadCXXThis();
3493       CGF.EmitStoreOfScalar(CXXThis, ThisLVal);
3494     }
3495     for (const LambdaCapture &LC : RD->captures()) {
3496       if (LC.getCaptureKind() != LCK_ByRef)
3497         continue;
3498       const ValueDecl *VD = LC.getCapturedVar();
3499       // FIXME: For now VD is always a VarDecl because OpenMP does not support
3500       //  capturing structured bindings in lambdas yet.
3501       if (!CS->capturesVariable(cast<VarDecl>(VD)))
3502         continue;
3503       auto It = Captures.find(VD);
3504       assert(It != Captures.end() && "Found lambda capture without field.");
3505       LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second);
3506       Address VDAddr = CGF.GetAddrOfLocalVar(cast<VarDecl>(VD));
3507       if (VD->getType().getCanonicalType()->isReferenceType())
3508         VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr,
3509                                                VD->getType().getCanonicalType())
3510                      .getAddress(CGF);
3511       CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal);
3512     }
3513   }
3514 }
3515 
3516 bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD,
3517                                                             LangAS &AS) {
3518   if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())
3519     return false;
3520   const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
3521   switch(A->getAllocatorType()) {
3522   case OMPAllocateDeclAttr::OMPNullMemAlloc:
3523   case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
3524   // Not supported, fallback to the default mem space.
3525   case OMPAllocateDeclAttr::OMPThreadMemAlloc:
3526   case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
3527   case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
3528   case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
3529   case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
3530     AS = LangAS::Default;
3531     return true;
3532   case OMPAllocateDeclAttr::OMPConstMemAlloc:
3533     AS = LangAS::cuda_constant;
3534     return true;
3535   case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
3536     AS = LangAS::cuda_shared;
3537     return true;
3538   case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
3539     llvm_unreachable("Expected predefined allocator for the variables with the "
3540                      "static storage.");
3541   }
3542   return false;
3543 }
3544 
3545 // Get current CudaArch and ignore any unknown values
3546 static CudaArch getCudaArch(CodeGenModule &CGM) {
3547   if (!CGM.getTarget().hasFeature("ptx"))
3548     return CudaArch::UNKNOWN;
3549   for (const auto &Feature : CGM.getTarget().getTargetOpts().FeatureMap) {
3550     if (Feature.getValue()) {
3551       CudaArch Arch = StringToCudaArch(Feature.getKey());
3552       if (Arch != CudaArch::UNKNOWN)
3553         return Arch;
3554     }
3555   }
3556   return CudaArch::UNKNOWN;
3557 }
3558 
3559 /// Check to see if target architecture supports unified addressing which is
3560 /// a restriction for OpenMP requires clause "unified_shared_memory".
3561 void CGOpenMPRuntimeGPU::processRequiresDirective(
3562     const OMPRequiresDecl *D) {
3563   for (const OMPClause *Clause : D->clauselists()) {
3564     if (Clause->getClauseKind() == OMPC_unified_shared_memory) {
3565       CudaArch Arch = getCudaArch(CGM);
3566       switch (Arch) {
3567       case CudaArch::SM_20:
3568       case CudaArch::SM_21:
3569       case CudaArch::SM_30:
3570       case CudaArch::SM_32:
3571       case CudaArch::SM_35:
3572       case CudaArch::SM_37:
3573       case CudaArch::SM_50:
3574       case CudaArch::SM_52:
3575       case CudaArch::SM_53: {
3576         SmallString<256> Buffer;
3577         llvm::raw_svector_ostream Out(Buffer);
3578         Out << "Target architecture " << CudaArchToString(Arch)
3579             << " does not support unified addressing";
3580         CGM.Error(Clause->getBeginLoc(), Out.str());
3581         return;
3582       }
3583       case CudaArch::SM_60:
3584       case CudaArch::SM_61:
3585       case CudaArch::SM_62:
3586       case CudaArch::SM_70:
3587       case CudaArch::SM_72:
3588       case CudaArch::SM_75:
3589       case CudaArch::SM_80:
3590       case CudaArch::SM_86:
3591       case CudaArch::SM_87:
3592       case CudaArch::SM_89:
3593       case CudaArch::SM_90:
3594       case CudaArch::GFX600:
3595       case CudaArch::GFX601:
3596       case CudaArch::GFX602:
3597       case CudaArch::GFX700:
3598       case CudaArch::GFX701:
3599       case CudaArch::GFX702:
3600       case CudaArch::GFX703:
3601       case CudaArch::GFX704:
3602       case CudaArch::GFX705:
3603       case CudaArch::GFX801:
3604       case CudaArch::GFX802:
3605       case CudaArch::GFX803:
3606       case CudaArch::GFX805:
3607       case CudaArch::GFX810:
3608       case CudaArch::GFX900:
3609       case CudaArch::GFX902:
3610       case CudaArch::GFX904:
3611       case CudaArch::GFX906:
3612       case CudaArch::GFX908:
3613       case CudaArch::GFX909:
3614       case CudaArch::GFX90a:
3615       case CudaArch::GFX90c:
3616       case CudaArch::GFX940:
3617       case CudaArch::GFX941:
3618       case CudaArch::GFX942:
3619       case CudaArch::GFX1010:
3620       case CudaArch::GFX1011:
3621       case CudaArch::GFX1012:
3622       case CudaArch::GFX1013:
3623       case CudaArch::GFX1030:
3624       case CudaArch::GFX1031:
3625       case CudaArch::GFX1032:
3626       case CudaArch::GFX1033:
3627       case CudaArch::GFX1034:
3628       case CudaArch::GFX1035:
3629       case CudaArch::GFX1036:
3630       case CudaArch::GFX1100:
3631       case CudaArch::GFX1101:
3632       case CudaArch::GFX1102:
3633       case CudaArch::GFX1103:
3634       case CudaArch::GFX1150:
3635       case CudaArch::GFX1151:
3636       case CudaArch::Generic:
3637       case CudaArch::UNUSED:
3638       case CudaArch::UNKNOWN:
3639         break;
3640       case CudaArch::LAST:
3641         llvm_unreachable("Unexpected Cuda arch.");
3642       }
3643     }
3644   }
3645   CGOpenMPRuntime::processRequiresDirective(D);
3646 }
3647 
3648 void CGOpenMPRuntimeGPU::clear() {
3649 
3650   if (!TeamsReductions.empty()) {
3651     ASTContext &C = CGM.getContext();
3652     RecordDecl *StaticRD = C.buildImplicitRecord(
3653         "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::TTK_Union);
3654     StaticRD->startDefinition();
3655     for (const RecordDecl *TeamReductionRec : TeamsReductions) {
3656       QualType RecTy = C.getRecordType(TeamReductionRec);
3657       auto *Field = FieldDecl::Create(
3658           C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy,
3659           C.getTrivialTypeSourceInfo(RecTy, SourceLocation()),
3660           /*BW=*/nullptr, /*Mutable=*/false,
3661           /*InitStyle=*/ICIS_NoInit);
3662       Field->setAccess(AS_public);
3663       StaticRD->addDecl(Field);
3664     }
3665     StaticRD->completeDefinition();
3666     QualType StaticTy = C.getRecordType(StaticRD);
3667     llvm::Type *LLVMReductionsBufferTy =
3668         CGM.getTypes().ConvertTypeForMem(StaticTy);
3669     // FIXME: nvlink does not handle weak linkage correctly (object with the
3670     // different size are reported as erroneous).
3671     // Restore CommonLinkage as soon as nvlink is fixed.
3672     auto *GV = new llvm::GlobalVariable(
3673         CGM.getModule(), LLVMReductionsBufferTy,
3674         /*isConstant=*/false, llvm::GlobalValue::InternalLinkage,
3675         llvm::Constant::getNullValue(LLVMReductionsBufferTy),
3676         "_openmp_teams_reductions_buffer_$_");
3677     KernelTeamsReductionPtr->setInitializer(
3678         llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV,
3679                                                              CGM.VoidPtrTy));
3680   }
3681   CGOpenMPRuntime::clear();
3682 }
3683 
3684 llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) {
3685   CGBuilderTy &Bld = CGF.Builder;
3686   llvm::Module *M = &CGF.CGM.getModule();
3687   const char *LocSize = "__kmpc_get_hardware_num_threads_in_block";
3688   llvm::Function *F = M->getFunction(LocSize);
3689   if (!F) {
3690     F = llvm::Function::Create(
3691         llvm::FunctionType::get(CGF.Int32Ty, std::nullopt, false),
3692         llvm::GlobalVariable::ExternalLinkage, LocSize, &CGF.CGM.getModule());
3693   }
3694   return Bld.CreateCall(F, std::nullopt, "nvptx_num_threads");
3695 }
3696 
3697 llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) {
3698   ArrayRef<llvm::Value *> Args{};
3699   return CGF.EmitRuntimeCall(
3700       OMPBuilder.getOrCreateRuntimeFunction(
3701           CGM.getModule(), OMPRTL___kmpc_get_hardware_thread_id_in_block),
3702       Args);
3703 }
3704 
3705 llvm::Value *CGOpenMPRuntimeGPU::getGPUWarpSize(CodeGenFunction &CGF) {
3706   ArrayRef<llvm::Value *> Args{};
3707   return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
3708                                  CGM.getModule(), OMPRTL___kmpc_get_warp_size),
3709                              Args);
3710 }
3711