xref: /freebsd/contrib/llvm-project/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
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 static const ValueDecl *getPrivateItem(const Expr *RefExpr) {
89   RefExpr = RefExpr->IgnoreParens();
90   if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) {
91     const Expr *Base = ASE->getBase()->IgnoreParenImpCasts();
92     while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
93       Base = TempASE->getBase()->IgnoreParenImpCasts();
94     RefExpr = Base;
95   } else if (auto *OASE = dyn_cast<ArraySectionExpr>(RefExpr)) {
96     const Expr *Base = OASE->getBase()->IgnoreParenImpCasts();
97     while (const auto *TempOASE = dyn_cast<ArraySectionExpr>(Base))
98       Base = TempOASE->getBase()->IgnoreParenImpCasts();
99     while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
100       Base = TempASE->getBase()->IgnoreParenImpCasts();
101     RefExpr = Base;
102   }
103   RefExpr = RefExpr->IgnoreParenImpCasts();
104   if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr))
105     return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl());
106   const auto *ME = cast<MemberExpr>(RefExpr);
107   return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl());
108 }
109 
110 static RecordDecl *buildRecordForGlobalizedVars(
111     ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls,
112     ArrayRef<const ValueDecl *> EscapedDeclsForTeams,
113     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
114         &MappedDeclsFields,
115     int BufSize) {
116   using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>;
117   if (EscapedDecls.empty() && EscapedDeclsForTeams.empty())
118     return nullptr;
119   SmallVector<VarsDataTy, 4> GlobalizedVars;
120   for (const ValueDecl *D : EscapedDecls)
121     GlobalizedVars.emplace_back(C.getDeclAlign(D), D);
122   for (const ValueDecl *D : EscapedDeclsForTeams)
123     GlobalizedVars.emplace_back(C.getDeclAlign(D), D);
124 
125   // Build struct _globalized_locals_ty {
126   //         /*  globalized vars  */[WarSize] align (decl_align)
127   //         /*  globalized vars  */ for EscapedDeclsForTeams
128   //       };
129   RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty");
130   GlobalizedRD->startDefinition();
131   llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped(
132       EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end());
133   for (const auto &Pair : GlobalizedVars) {
134     const ValueDecl *VD = Pair.second;
135     QualType Type = VD->getType();
136     if (Type->isLValueReferenceType())
137       Type = C.getPointerType(Type.getNonReferenceType());
138     else
139       Type = Type.getNonReferenceType();
140     SourceLocation Loc = VD->getLocation();
141     FieldDecl *Field;
142     if (SingleEscaped.count(VD)) {
143       Field = FieldDecl::Create(
144           C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
145           C.getTrivialTypeSourceInfo(Type, SourceLocation()),
146           /*BW=*/nullptr, /*Mutable=*/false,
147           /*InitStyle=*/ICIS_NoInit);
148       Field->setAccess(AS_public);
149       if (VD->hasAttrs()) {
150         for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
151              E(VD->getAttrs().end());
152              I != E; ++I)
153           Field->addAttr(*I);
154       }
155     } else {
156       if (BufSize > 1) {
157         llvm::APInt ArraySize(32, BufSize);
158         Type = C.getConstantArrayType(Type, ArraySize, nullptr,
159                                       ArraySizeModifier::Normal, 0);
160       }
161       Field = FieldDecl::Create(
162           C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
163           C.getTrivialTypeSourceInfo(Type, SourceLocation()),
164           /*BW=*/nullptr, /*Mutable=*/false,
165           /*InitStyle=*/ICIS_NoInit);
166       Field->setAccess(AS_public);
167       llvm::APInt Align(32, Pair.first.getQuantity());
168       Field->addAttr(AlignedAttr::CreateImplicit(
169           C, /*IsAlignmentExpr=*/true,
170           IntegerLiteral::Create(C, Align,
171                                  C.getIntTypeForBitwidth(32, /*Signed=*/0),
172                                  SourceLocation()),
173           {}, AlignedAttr::GNU_aligned));
174     }
175     GlobalizedRD->addDecl(Field);
176     MappedDeclsFields.try_emplace(VD, Field);
177   }
178   GlobalizedRD->completeDefinition();
179   return GlobalizedRD;
180 }
181 
182 /// Get the list of variables that can escape their declaration context.
183 class CheckVarsEscapingDeclContext final
184     : public ConstStmtVisitor<CheckVarsEscapingDeclContext> {
185   CodeGenFunction &CGF;
186   llvm::SetVector<const ValueDecl *> EscapedDecls;
187   llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls;
188   llvm::SetVector<const ValueDecl *> DelayedVariableLengthDecls;
189   llvm::SmallPtrSet<const Decl *, 4> EscapedParameters;
190   RecordDecl *GlobalizedRD = nullptr;
191   llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
192   bool AllEscaped = false;
193   bool IsForCombinedParallelRegion = false;
194 
195   void markAsEscaped(const ValueDecl *VD) {
196     // Do not globalize declare target variables.
197     if (!isa<VarDecl>(VD) ||
198         OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD))
199       return;
200     VD = cast<ValueDecl>(VD->getCanonicalDecl());
201     // Use user-specified allocation.
202     if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>())
203       return;
204     // Variables captured by value must be globalized.
205     bool IsCaptured = false;
206     if (auto *CSI = CGF.CapturedStmtInfo) {
207       if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) {
208         // Check if need to capture the variable that was already captured by
209         // value in the outer region.
210         IsCaptured = true;
211         if (!IsForCombinedParallelRegion) {
212           if (!FD->hasAttrs())
213             return;
214           const auto *Attr = FD->getAttr<OMPCaptureKindAttr>();
215           if (!Attr)
216             return;
217           if (((Attr->getCaptureKind() != OMPC_map) &&
218                !isOpenMPPrivate(Attr->getCaptureKind())) ||
219               ((Attr->getCaptureKind() == OMPC_map) &&
220                !FD->getType()->isAnyPointerType()))
221             return;
222         }
223         if (!FD->getType()->isReferenceType()) {
224           assert(!VD->getType()->isVariablyModifiedType() &&
225                  "Parameter captured by value with variably modified type");
226           EscapedParameters.insert(VD);
227         } else if (!IsForCombinedParallelRegion) {
228           return;
229         }
230       }
231     }
232     if ((!CGF.CapturedStmtInfo ||
233          (IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) &&
234         VD->getType()->isReferenceType())
235       // Do not globalize variables with reference type.
236       return;
237     if (VD->getType()->isVariablyModifiedType()) {
238       // If not captured at the target region level then mark the escaped
239       // variable as delayed.
240       if (IsCaptured)
241         EscapedVariableLengthDecls.insert(VD);
242       else
243         DelayedVariableLengthDecls.insert(VD);
244     } else
245       EscapedDecls.insert(VD);
246   }
247 
248   void VisitValueDecl(const ValueDecl *VD) {
249     if (VD->getType()->isLValueReferenceType())
250       markAsEscaped(VD);
251     if (const auto *VarD = dyn_cast<VarDecl>(VD)) {
252       if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) {
253         const bool SavedAllEscaped = AllEscaped;
254         AllEscaped = VD->getType()->isLValueReferenceType();
255         Visit(VarD->getInit());
256         AllEscaped = SavedAllEscaped;
257       }
258     }
259   }
260   void VisitOpenMPCapturedStmt(const CapturedStmt *S,
261                                ArrayRef<OMPClause *> Clauses,
262                                bool IsCombinedParallelRegion) {
263     if (!S)
264       return;
265     for (const CapturedStmt::Capture &C : S->captures()) {
266       if (C.capturesVariable() && !C.capturesVariableByCopy()) {
267         const ValueDecl *VD = C.getCapturedVar();
268         bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion;
269         if (IsCombinedParallelRegion) {
270           // Check if the variable is privatized in the combined construct and
271           // those private copies must be shared in the inner parallel
272           // directive.
273           IsForCombinedParallelRegion = false;
274           for (const OMPClause *C : Clauses) {
275             if (!isOpenMPPrivate(C->getClauseKind()) ||
276                 C->getClauseKind() == OMPC_reduction ||
277                 C->getClauseKind() == OMPC_linear ||
278                 C->getClauseKind() == OMPC_private)
279               continue;
280             ArrayRef<const Expr *> Vars;
281             if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C))
282               Vars = PC->getVarRefs();
283             else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C))
284               Vars = PC->getVarRefs();
285             else
286               llvm_unreachable("Unexpected clause.");
287             for (const auto *E : Vars) {
288               const Decl *D =
289                   cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl();
290               if (D == VD->getCanonicalDecl()) {
291                 IsForCombinedParallelRegion = true;
292                 break;
293               }
294             }
295             if (IsForCombinedParallelRegion)
296               break;
297           }
298         }
299         markAsEscaped(VD);
300         if (isa<OMPCapturedExprDecl>(VD))
301           VisitValueDecl(VD);
302         IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion;
303       }
304     }
305   }
306 
307   void buildRecordForGlobalizedVars(bool IsInTTDRegion) {
308     assert(!GlobalizedRD &&
309            "Record for globalized variables is built already.");
310     ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams;
311     unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size;
312     if (IsInTTDRegion)
313       EscapedDeclsForTeams = EscapedDecls.getArrayRef();
314     else
315       EscapedDeclsForParallel = EscapedDecls.getArrayRef();
316     GlobalizedRD = ::buildRecordForGlobalizedVars(
317         CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams,
318         MappedDeclsFields, WarpSize);
319   }
320 
321 public:
322   CheckVarsEscapingDeclContext(CodeGenFunction &CGF,
323                                ArrayRef<const ValueDecl *> TeamsReductions)
324       : CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) {
325   }
326   virtual ~CheckVarsEscapingDeclContext() = default;
327   void VisitDeclStmt(const DeclStmt *S) {
328     if (!S)
329       return;
330     for (const Decl *D : S->decls())
331       if (const auto *VD = dyn_cast_or_null<ValueDecl>(D))
332         VisitValueDecl(VD);
333   }
334   void VisitOMPExecutableDirective(const OMPExecutableDirective *D) {
335     if (!D)
336       return;
337     if (!D->hasAssociatedStmt())
338       return;
339     if (const auto *S =
340             dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) {
341       // Do not analyze directives that do not actually require capturing,
342       // like `omp for` or `omp simd` directives.
343       llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
344       getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind());
345       if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) {
346         VisitStmt(S->getCapturedStmt());
347         return;
348       }
349       VisitOpenMPCapturedStmt(
350           S, D->clauses(),
351           CaptureRegions.back() == OMPD_parallel &&
352               isOpenMPDistributeDirective(D->getDirectiveKind()));
353     }
354   }
355   void VisitCapturedStmt(const CapturedStmt *S) {
356     if (!S)
357       return;
358     for (const CapturedStmt::Capture &C : S->captures()) {
359       if (C.capturesVariable() && !C.capturesVariableByCopy()) {
360         const ValueDecl *VD = C.getCapturedVar();
361         markAsEscaped(VD);
362         if (isa<OMPCapturedExprDecl>(VD))
363           VisitValueDecl(VD);
364       }
365     }
366   }
367   void VisitLambdaExpr(const LambdaExpr *E) {
368     if (!E)
369       return;
370     for (const LambdaCapture &C : E->captures()) {
371       if (C.capturesVariable()) {
372         if (C.getCaptureKind() == LCK_ByRef) {
373           const ValueDecl *VD = C.getCapturedVar();
374           markAsEscaped(VD);
375           if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD))
376             VisitValueDecl(VD);
377         }
378       }
379     }
380   }
381   void VisitBlockExpr(const BlockExpr *E) {
382     if (!E)
383       return;
384     for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) {
385       if (C.isByRef()) {
386         const VarDecl *VD = C.getVariable();
387         markAsEscaped(VD);
388         if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture())
389           VisitValueDecl(VD);
390       }
391     }
392   }
393   void VisitCallExpr(const CallExpr *E) {
394     if (!E)
395       return;
396     for (const Expr *Arg : E->arguments()) {
397       if (!Arg)
398         continue;
399       if (Arg->isLValue()) {
400         const bool SavedAllEscaped = AllEscaped;
401         AllEscaped = true;
402         Visit(Arg);
403         AllEscaped = SavedAllEscaped;
404       } else {
405         Visit(Arg);
406       }
407     }
408     Visit(E->getCallee());
409   }
410   void VisitDeclRefExpr(const DeclRefExpr *E) {
411     if (!E)
412       return;
413     const ValueDecl *VD = E->getDecl();
414     if (AllEscaped)
415       markAsEscaped(VD);
416     if (isa<OMPCapturedExprDecl>(VD))
417       VisitValueDecl(VD);
418     else if (VD->isInitCapture())
419       VisitValueDecl(VD);
420   }
421   void VisitUnaryOperator(const UnaryOperator *E) {
422     if (!E)
423       return;
424     if (E->getOpcode() == UO_AddrOf) {
425       const bool SavedAllEscaped = AllEscaped;
426       AllEscaped = true;
427       Visit(E->getSubExpr());
428       AllEscaped = SavedAllEscaped;
429     } else {
430       Visit(E->getSubExpr());
431     }
432   }
433   void VisitImplicitCastExpr(const ImplicitCastExpr *E) {
434     if (!E)
435       return;
436     if (E->getCastKind() == CK_ArrayToPointerDecay) {
437       const bool SavedAllEscaped = AllEscaped;
438       AllEscaped = true;
439       Visit(E->getSubExpr());
440       AllEscaped = SavedAllEscaped;
441     } else {
442       Visit(E->getSubExpr());
443     }
444   }
445   void VisitExpr(const Expr *E) {
446     if (!E)
447       return;
448     bool SavedAllEscaped = AllEscaped;
449     if (!E->isLValue())
450       AllEscaped = false;
451     for (const Stmt *Child : E->children())
452       if (Child)
453         Visit(Child);
454     AllEscaped = SavedAllEscaped;
455   }
456   void VisitStmt(const Stmt *S) {
457     if (!S)
458       return;
459     for (const Stmt *Child : S->children())
460       if (Child)
461         Visit(Child);
462   }
463 
464   /// Returns the record that handles all the escaped local variables and used
465   /// instead of their original storage.
466   const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) {
467     if (!GlobalizedRD)
468       buildRecordForGlobalizedVars(IsInTTDRegion);
469     return GlobalizedRD;
470   }
471 
472   /// Returns the field in the globalized record for the escaped variable.
473   const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const {
474     assert(GlobalizedRD &&
475            "Record for globalized variables must be generated already.");
476     return MappedDeclsFields.lookup(VD);
477   }
478 
479   /// Returns the list of the escaped local variables/parameters.
480   ArrayRef<const ValueDecl *> getEscapedDecls() const {
481     return EscapedDecls.getArrayRef();
482   }
483 
484   /// Checks if the escaped local variable is actually a parameter passed by
485   /// value.
486   const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const {
487     return EscapedParameters;
488   }
489 
490   /// Returns the list of the escaped variables with the variably modified
491   /// types.
492   ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const {
493     return EscapedVariableLengthDecls.getArrayRef();
494   }
495 
496   /// Returns the list of the delayed variables with the variably modified
497   /// types.
498   ArrayRef<const ValueDecl *> getDelayedVariableLengthDecls() const {
499     return DelayedVariableLengthDecls.getArrayRef();
500   }
501 };
502 } // anonymous namespace
503 
504 CGOpenMPRuntimeGPU::ExecutionMode
505 CGOpenMPRuntimeGPU::getExecutionMode() const {
506   return CurrentExecutionMode;
507 }
508 
509 CGOpenMPRuntimeGPU::DataSharingMode
510 CGOpenMPRuntimeGPU::getDataSharingMode() const {
511   return CurrentDataSharingMode;
512 }
513 
514 /// Check for inner (nested) SPMD construct, if any
515 static bool hasNestedSPMDDirective(ASTContext &Ctx,
516                                    const OMPExecutableDirective &D) {
517   const auto *CS = D.getInnermostCapturedStmt();
518   const auto *Body =
519       CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
520   const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
521 
522   if (const auto *NestedDir =
523           dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
524     OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
525     switch (D.getDirectiveKind()) {
526     case OMPD_target:
527       if (isOpenMPParallelDirective(DKind))
528         return true;
529       if (DKind == OMPD_teams) {
530         Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
531             /*IgnoreCaptured=*/true);
532         if (!Body)
533           return false;
534         ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
535         if (const auto *NND =
536                 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
537           DKind = NND->getDirectiveKind();
538           if (isOpenMPParallelDirective(DKind))
539             return true;
540         }
541       }
542       return false;
543     case OMPD_target_teams:
544       return isOpenMPParallelDirective(DKind);
545     case OMPD_target_simd:
546     case OMPD_target_parallel:
547     case OMPD_target_parallel_for:
548     case OMPD_target_parallel_for_simd:
549     case OMPD_target_teams_distribute:
550     case OMPD_target_teams_distribute_simd:
551     case OMPD_target_teams_distribute_parallel_for:
552     case OMPD_target_teams_distribute_parallel_for_simd:
553     case OMPD_parallel:
554     case OMPD_for:
555     case OMPD_parallel_for:
556     case OMPD_parallel_master:
557     case OMPD_parallel_sections:
558     case OMPD_for_simd:
559     case OMPD_parallel_for_simd:
560     case OMPD_cancel:
561     case OMPD_cancellation_point:
562     case OMPD_ordered:
563     case OMPD_threadprivate:
564     case OMPD_allocate:
565     case OMPD_task:
566     case OMPD_simd:
567     case OMPD_sections:
568     case OMPD_section:
569     case OMPD_single:
570     case OMPD_master:
571     case OMPD_critical:
572     case OMPD_taskyield:
573     case OMPD_barrier:
574     case OMPD_taskwait:
575     case OMPD_taskgroup:
576     case OMPD_atomic:
577     case OMPD_flush:
578     case OMPD_depobj:
579     case OMPD_scan:
580     case OMPD_teams:
581     case OMPD_target_data:
582     case OMPD_target_exit_data:
583     case OMPD_target_enter_data:
584     case OMPD_distribute:
585     case OMPD_distribute_simd:
586     case OMPD_distribute_parallel_for:
587     case OMPD_distribute_parallel_for_simd:
588     case OMPD_teams_distribute:
589     case OMPD_teams_distribute_simd:
590     case OMPD_teams_distribute_parallel_for:
591     case OMPD_teams_distribute_parallel_for_simd:
592     case OMPD_target_update:
593     case OMPD_declare_simd:
594     case OMPD_declare_variant:
595     case OMPD_begin_declare_variant:
596     case OMPD_end_declare_variant:
597     case OMPD_declare_target:
598     case OMPD_end_declare_target:
599     case OMPD_declare_reduction:
600     case OMPD_declare_mapper:
601     case OMPD_taskloop:
602     case OMPD_taskloop_simd:
603     case OMPD_master_taskloop:
604     case OMPD_master_taskloop_simd:
605     case OMPD_parallel_master_taskloop:
606     case OMPD_parallel_master_taskloop_simd:
607     case OMPD_requires:
608     case OMPD_unknown:
609     default:
610       llvm_unreachable("Unexpected directive.");
611     }
612   }
613 
614   return false;
615 }
616 
617 static bool supportsSPMDExecutionMode(ASTContext &Ctx,
618                                       const OMPExecutableDirective &D) {
619   OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
620   switch (DirectiveKind) {
621   case OMPD_target:
622   case OMPD_target_teams:
623     return hasNestedSPMDDirective(Ctx, D);
624   case OMPD_target_parallel_loop:
625   case OMPD_target_parallel:
626   case OMPD_target_parallel_for:
627   case OMPD_target_parallel_for_simd:
628   case OMPD_target_teams_distribute_parallel_for:
629   case OMPD_target_teams_distribute_parallel_for_simd:
630   case OMPD_target_simd:
631   case OMPD_target_teams_distribute_simd:
632     return true;
633   case OMPD_target_teams_distribute:
634     return false;
635   case OMPD_target_teams_loop:
636     // Whether this is true or not depends on how the directive will
637     // eventually be emitted.
638     if (auto *TTLD = dyn_cast<OMPTargetTeamsGenericLoopDirective>(&D))
639       return TTLD->canBeParallelFor();
640     return false;
641   case OMPD_parallel:
642   case OMPD_for:
643   case OMPD_parallel_for:
644   case OMPD_parallel_master:
645   case OMPD_parallel_sections:
646   case OMPD_for_simd:
647   case OMPD_parallel_for_simd:
648   case OMPD_cancel:
649   case OMPD_cancellation_point:
650   case OMPD_ordered:
651   case OMPD_threadprivate:
652   case OMPD_allocate:
653   case OMPD_task:
654   case OMPD_simd:
655   case OMPD_sections:
656   case OMPD_section:
657   case OMPD_single:
658   case OMPD_master:
659   case OMPD_critical:
660   case OMPD_taskyield:
661   case OMPD_barrier:
662   case OMPD_taskwait:
663   case OMPD_taskgroup:
664   case OMPD_atomic:
665   case OMPD_flush:
666   case OMPD_depobj:
667   case OMPD_scan:
668   case OMPD_teams:
669   case OMPD_target_data:
670   case OMPD_target_exit_data:
671   case OMPD_target_enter_data:
672   case OMPD_distribute:
673   case OMPD_distribute_simd:
674   case OMPD_distribute_parallel_for:
675   case OMPD_distribute_parallel_for_simd:
676   case OMPD_teams_distribute:
677   case OMPD_teams_distribute_simd:
678   case OMPD_teams_distribute_parallel_for:
679   case OMPD_teams_distribute_parallel_for_simd:
680   case OMPD_target_update:
681   case OMPD_declare_simd:
682   case OMPD_declare_variant:
683   case OMPD_begin_declare_variant:
684   case OMPD_end_declare_variant:
685   case OMPD_declare_target:
686   case OMPD_end_declare_target:
687   case OMPD_declare_reduction:
688   case OMPD_declare_mapper:
689   case OMPD_taskloop:
690   case OMPD_taskloop_simd:
691   case OMPD_master_taskloop:
692   case OMPD_master_taskloop_simd:
693   case OMPD_parallel_master_taskloop:
694   case OMPD_parallel_master_taskloop_simd:
695   case OMPD_requires:
696   case OMPD_unknown:
697   default:
698     break;
699   }
700   llvm_unreachable(
701       "Unknown programming model for OpenMP directive on NVPTX target.");
702 }
703 
704 void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D,
705                                              StringRef ParentName,
706                                              llvm::Function *&OutlinedFn,
707                                              llvm::Constant *&OutlinedFnID,
708                                              bool IsOffloadEntry,
709                                              const RegionCodeGenTy &CodeGen) {
710   ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_NonSPMD);
711   EntryFunctionState EST;
712   WrapperFunctionsMap.clear();
713 
714   [[maybe_unused]] bool IsBareKernel = D.getSingleClause<OMPXBareClause>();
715   assert(!IsBareKernel && "bare kernel should not be at generic mode");
716 
717   // Emit target region as a standalone region.
718   class NVPTXPrePostActionTy : public PrePostActionTy {
719     CGOpenMPRuntimeGPU::EntryFunctionState &EST;
720     const OMPExecutableDirective &D;
721 
722   public:
723     NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST,
724                          const OMPExecutableDirective &D)
725         : EST(EST), D(D) {}
726     void Enter(CodeGenFunction &CGF) override {
727       auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
728       RT.emitKernelInit(D, CGF, EST, /* IsSPMD */ false);
729       // Skip target region initialization.
730       RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
731     }
732     void Exit(CodeGenFunction &CGF) override {
733       auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
734       RT.clearLocThreadIdInsertPt(CGF);
735       RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ false);
736     }
737   } Action(EST, D);
738   CodeGen.setAction(Action);
739   IsInTTDRegion = true;
740   emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
741                                    IsOffloadEntry, CodeGen);
742   IsInTTDRegion = false;
743 }
744 
745 void CGOpenMPRuntimeGPU::emitKernelInit(const OMPExecutableDirective &D,
746                                         CodeGenFunction &CGF,
747                                         EntryFunctionState &EST, bool IsSPMD) {
748   int32_t MinThreadsVal = 1, MaxThreadsVal = -1, MinTeamsVal = 1,
749           MaxTeamsVal = -1;
750   computeMinAndMaxThreadsAndTeams(D, CGF, MinThreadsVal, MaxThreadsVal,
751                                   MinTeamsVal, MaxTeamsVal);
752 
753   CGBuilderTy &Bld = CGF.Builder;
754   Bld.restoreIP(OMPBuilder.createTargetInit(
755       Bld, IsSPMD, MinThreadsVal, MaxThreadsVal, MinTeamsVal, MaxTeamsVal));
756   if (!IsSPMD)
757     emitGenericVarsProlog(CGF, EST.Loc);
758 }
759 
760 void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF,
761                                           EntryFunctionState &EST,
762                                           bool IsSPMD) {
763   if (!IsSPMD)
764     emitGenericVarsEpilog(CGF);
765 
766   // This is temporary until we remove the fixed sized buffer.
767   ASTContext &C = CGM.getContext();
768   RecordDecl *StaticRD = C.buildImplicitRecord(
769       "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::Union);
770   StaticRD->startDefinition();
771   for (const RecordDecl *TeamReductionRec : TeamsReductions) {
772     QualType RecTy = C.getRecordType(TeamReductionRec);
773     auto *Field = FieldDecl::Create(
774         C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy,
775         C.getTrivialTypeSourceInfo(RecTy, SourceLocation()),
776         /*BW=*/nullptr, /*Mutable=*/false,
777         /*InitStyle=*/ICIS_NoInit);
778     Field->setAccess(AS_public);
779     StaticRD->addDecl(Field);
780   }
781   StaticRD->completeDefinition();
782   QualType StaticTy = C.getRecordType(StaticRD);
783   llvm::Type *LLVMReductionsBufferTy =
784       CGM.getTypes().ConvertTypeForMem(StaticTy);
785   const auto &DL = CGM.getModule().getDataLayout();
786   uint64_t ReductionDataSize =
787       TeamsReductions.empty()
788           ? 0
789           : DL.getTypeAllocSize(LLVMReductionsBufferTy).getFixedValue();
790   CGBuilderTy &Bld = CGF.Builder;
791   OMPBuilder.createTargetDeinit(Bld, ReductionDataSize,
792                                 C.getLangOpts().OpenMPCUDAReductionBufNum);
793   TeamsReductions.clear();
794 }
795 
796 void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D,
797                                           StringRef ParentName,
798                                           llvm::Function *&OutlinedFn,
799                                           llvm::Constant *&OutlinedFnID,
800                                           bool IsOffloadEntry,
801                                           const RegionCodeGenTy &CodeGen) {
802   ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_SPMD);
803   EntryFunctionState EST;
804 
805   bool IsBareKernel = D.getSingleClause<OMPXBareClause>();
806 
807   // Emit target region as a standalone region.
808   class NVPTXPrePostActionTy : public PrePostActionTy {
809     CGOpenMPRuntimeGPU &RT;
810     CGOpenMPRuntimeGPU::EntryFunctionState &EST;
811     bool IsBareKernel;
812     DataSharingMode Mode;
813     const OMPExecutableDirective &D;
814 
815   public:
816     NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT,
817                          CGOpenMPRuntimeGPU::EntryFunctionState &EST,
818                          bool IsBareKernel, const OMPExecutableDirective &D)
819         : RT(RT), EST(EST), IsBareKernel(IsBareKernel),
820           Mode(RT.CurrentDataSharingMode), D(D) {}
821     void Enter(CodeGenFunction &CGF) override {
822       if (IsBareKernel) {
823         RT.CurrentDataSharingMode = DataSharingMode::DS_CUDA;
824         return;
825       }
826       RT.emitKernelInit(D, CGF, EST, /* IsSPMD */ true);
827       // Skip target region initialization.
828       RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
829     }
830     void Exit(CodeGenFunction &CGF) override {
831       if (IsBareKernel) {
832         RT.CurrentDataSharingMode = Mode;
833         return;
834       }
835       RT.clearLocThreadIdInsertPt(CGF);
836       RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true);
837     }
838   } Action(*this, EST, IsBareKernel, D);
839   CodeGen.setAction(Action);
840   IsInTTDRegion = true;
841   emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
842                                    IsOffloadEntry, CodeGen);
843   IsInTTDRegion = false;
844 }
845 
846 void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction(
847     const OMPExecutableDirective &D, StringRef ParentName,
848     llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
849     bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
850   if (!IsOffloadEntry) // Nothing to do.
851     return;
852 
853   assert(!ParentName.empty() && "Invalid target region parent name!");
854 
855   bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D);
856   bool IsBareKernel = D.getSingleClause<OMPXBareClause>();
857   if (Mode || IsBareKernel)
858     emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
859                    CodeGen);
860   else
861     emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
862                       CodeGen);
863 }
864 
865 CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM)
866     : CGOpenMPRuntime(CGM) {
867   llvm::OpenMPIRBuilderConfig Config(
868       CGM.getLangOpts().OpenMPIsTargetDevice, isGPU(),
869       CGM.getLangOpts().OpenMPOffloadMandatory,
870       /*HasRequiresReverseOffload*/ false, /*HasRequiresUnifiedAddress*/ false,
871       hasRequiresUnifiedSharedMemory(), /*HasRequiresDynamicAllocators*/ false);
872   OMPBuilder.setConfig(Config);
873 
874   if (!CGM.getLangOpts().OpenMPIsTargetDevice)
875     llvm_unreachable("OpenMP can only handle device code.");
876 
877   if (CGM.getLangOpts().OpenMPCUDAMode)
878     CurrentDataSharingMode = CGOpenMPRuntimeGPU::DS_CUDA;
879 
880   llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder();
881   if (CGM.getLangOpts().NoGPULib || CGM.getLangOpts().OMPHostIRFile.empty())
882     return;
883 
884   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTargetDebug,
885                               "__omp_rtl_debug_kind");
886   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTeamSubscription,
887                               "__omp_rtl_assume_teams_oversubscription");
888   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPThreadSubscription,
889                               "__omp_rtl_assume_threads_oversubscription");
890   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoThreadState,
891                               "__omp_rtl_assume_no_thread_state");
892   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoNestedParallelism,
893                               "__omp_rtl_assume_no_nested_parallelism");
894 }
895 
896 void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF,
897                                               ProcBindKind ProcBind,
898                                               SourceLocation Loc) {
899   // Nothing to do.
900 }
901 
902 void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF,
903                                                 llvm::Value *NumThreads,
904                                                 SourceLocation Loc) {
905   // Nothing to do.
906 }
907 
908 void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF,
909                                               const Expr *NumTeams,
910                                               const Expr *ThreadLimit,
911                                               SourceLocation Loc) {}
912 
913 llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction(
914     CodeGenFunction &CGF, const OMPExecutableDirective &D,
915     const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind,
916     const RegionCodeGenTy &CodeGen) {
917   // Emit target region as a standalone region.
918   bool PrevIsInTTDRegion = IsInTTDRegion;
919   IsInTTDRegion = false;
920   auto *OutlinedFun =
921       cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction(
922           CGF, D, ThreadIDVar, InnermostKind, CodeGen));
923   IsInTTDRegion = PrevIsInTTDRegion;
924   if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) {
925     llvm::Function *WrapperFun =
926         createParallelDataSharingWrapper(OutlinedFun, D);
927     WrapperFunctionsMap[OutlinedFun] = WrapperFun;
928   }
929 
930   return OutlinedFun;
931 }
932 
933 /// Get list of lastprivate variables from the teams distribute ... or
934 /// teams {distribute ...} directives.
935 static void
936 getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D,
937                              llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
938   assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&
939          "expected teams directive.");
940   const OMPExecutableDirective *Dir = &D;
941   if (!isOpenMPDistributeDirective(D.getDirectiveKind())) {
942     if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild(
943             Ctx,
944             D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(
945                 /*IgnoreCaptured=*/true))) {
946       Dir = dyn_cast_or_null<OMPExecutableDirective>(S);
947       if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind()))
948         Dir = nullptr;
949     }
950   }
951   if (!Dir)
952     return;
953   for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) {
954     for (const Expr *E : C->getVarRefs())
955       Vars.push_back(getPrivateItem(E));
956   }
957 }
958 
959 /// Get list of reduction variables from the teams ... directives.
960 static void
961 getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D,
962                       llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
963   assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&
964          "expected teams directive.");
965   for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
966     for (const Expr *E : C->privates())
967       Vars.push_back(getPrivateItem(E));
968   }
969 }
970 
971 llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction(
972     CodeGenFunction &CGF, const OMPExecutableDirective &D,
973     const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind,
974     const RegionCodeGenTy &CodeGen) {
975   SourceLocation Loc = D.getBeginLoc();
976 
977   const RecordDecl *GlobalizedRD = nullptr;
978   llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions;
979   llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
980   unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size;
981   // Globalize team reductions variable unconditionally in all modes.
982   if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
983     getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions);
984   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
985     getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions);
986     if (!LastPrivatesReductions.empty()) {
987       GlobalizedRD = ::buildRecordForGlobalizedVars(
988           CGM.getContext(), std::nullopt, LastPrivatesReductions,
989           MappedDeclsFields, WarpSize);
990     }
991   } else if (!LastPrivatesReductions.empty()) {
992     assert(!TeamAndReductions.first &&
993            "Previous team declaration is not expected.");
994     TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl();
995     std::swap(TeamAndReductions.second, LastPrivatesReductions);
996   }
997 
998   // Emit target region as a standalone region.
999   class NVPTXPrePostActionTy : public PrePostActionTy {
1000     SourceLocation &Loc;
1001     const RecordDecl *GlobalizedRD;
1002     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
1003         &MappedDeclsFields;
1004 
1005   public:
1006     NVPTXPrePostActionTy(
1007         SourceLocation &Loc, const RecordDecl *GlobalizedRD,
1008         llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
1009             &MappedDeclsFields)
1010         : Loc(Loc), GlobalizedRD(GlobalizedRD),
1011           MappedDeclsFields(MappedDeclsFields) {}
1012     void Enter(CodeGenFunction &CGF) override {
1013       auto &Rt =
1014           static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1015       if (GlobalizedRD) {
1016         auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
1017         I->getSecond().MappedParams =
1018             std::make_unique<CodeGenFunction::OMPMapVars>();
1019         DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
1020         for (const auto &Pair : MappedDeclsFields) {
1021           assert(Pair.getFirst()->isCanonicalDecl() &&
1022                  "Expected canonical declaration");
1023           Data.insert(std::make_pair(Pair.getFirst(), MappedVarData()));
1024         }
1025       }
1026       Rt.emitGenericVarsProlog(CGF, Loc);
1027     }
1028     void Exit(CodeGenFunction &CGF) override {
1029       static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
1030           .emitGenericVarsEpilog(CGF);
1031     }
1032   } Action(Loc, GlobalizedRD, MappedDeclsFields);
1033   CodeGen.setAction(Action);
1034   llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction(
1035       CGF, D, ThreadIDVar, InnermostKind, CodeGen);
1036 
1037   return OutlinedFun;
1038 }
1039 
1040 void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF,
1041                                                SourceLocation Loc) {
1042   if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic)
1043     return;
1044 
1045   CGBuilderTy &Bld = CGF.Builder;
1046 
1047   const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
1048   if (I == FunctionGlobalizedDecls.end())
1049     return;
1050 
1051   for (auto &Rec : I->getSecond().LocalVarData) {
1052     const auto *VD = cast<VarDecl>(Rec.first);
1053     bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first);
1054     QualType VarTy = VD->getType();
1055 
1056     // Get the local allocation of a firstprivate variable before sharing
1057     llvm::Value *ParValue;
1058     if (EscapedParam) {
1059       LValue ParLVal =
1060           CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
1061       ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc);
1062     }
1063 
1064     // Allocate space for the variable to be globalized
1065     llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())};
1066     llvm::CallBase *VoidPtr =
1067         CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1068                                 CGM.getModule(), OMPRTL___kmpc_alloc_shared),
1069                             AllocArgs, VD->getName());
1070     // FIXME: We should use the variables actual alignment as an argument.
1071     VoidPtr->addRetAttr(llvm::Attribute::get(
1072         CGM.getLLVMContext(), llvm::Attribute::Alignment,
1073         CGM.getContext().getTargetInfo().getNewAlign() / 8));
1074 
1075     // Cast the void pointer and get the address of the globalized variable.
1076     llvm::PointerType *VarPtrTy = CGF.ConvertTypeForMem(VarTy)->getPointerTo();
1077     llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
1078         VoidPtr, VarPtrTy, VD->getName() + "_on_stack");
1079     LValue VarAddr =
1080         CGF.MakeNaturalAlignPointeeRawAddrLValue(CastedVoidPtr, VarTy);
1081     Rec.second.PrivateAddr = VarAddr.getAddress();
1082     Rec.second.GlobalizedVal = VoidPtr;
1083 
1084     // Assign the local allocation to the newly globalized location.
1085     if (EscapedParam) {
1086       CGF.EmitStoreOfScalar(ParValue, VarAddr);
1087       I->getSecond().MappedParams->setVarAddr(CGF, VD, VarAddr.getAddress());
1088     }
1089     if (auto *DI = CGF.getDebugInfo())
1090       VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(VD->getLocation()));
1091   }
1092 
1093   for (const auto *ValueD : I->getSecond().EscapedVariableLengthDecls) {
1094     const auto *VD = cast<VarDecl>(ValueD);
1095     std::pair<llvm::Value *, llvm::Value *> AddrSizePair =
1096         getKmpcAllocShared(CGF, VD);
1097     I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(AddrSizePair);
1098     LValue Base = CGF.MakeAddrLValue(AddrSizePair.first, VD->getType(),
1099                                      CGM.getContext().getDeclAlign(VD),
1100                                      AlignmentSource::Decl);
1101     I->getSecond().MappedParams->setVarAddr(CGF, VD, Base.getAddress());
1102   }
1103   I->getSecond().MappedParams->apply(CGF);
1104 }
1105 
1106 bool CGOpenMPRuntimeGPU::isDelayedVariableLengthDecl(CodeGenFunction &CGF,
1107                                                      const VarDecl *VD) const {
1108   const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
1109   if (I == FunctionGlobalizedDecls.end())
1110     return false;
1111 
1112   // Check variable declaration is delayed:
1113   return llvm::is_contained(I->getSecond().DelayedVariableLengthDecls, VD);
1114 }
1115 
1116 std::pair<llvm::Value *, llvm::Value *>
1117 CGOpenMPRuntimeGPU::getKmpcAllocShared(CodeGenFunction &CGF,
1118                                        const VarDecl *VD) {
1119   CGBuilderTy &Bld = CGF.Builder;
1120 
1121   // Compute size and alignment.
1122   llvm::Value *Size = CGF.getTypeSize(VD->getType());
1123   CharUnits Align = CGM.getContext().getDeclAlign(VD);
1124   Size = Bld.CreateNUWAdd(
1125       Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1));
1126   llvm::Value *AlignVal =
1127       llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity());
1128   Size = Bld.CreateUDiv(Size, AlignVal);
1129   Size = Bld.CreateNUWMul(Size, AlignVal);
1130 
1131   // Allocate space for this VLA object to be globalized.
1132   llvm::Value *AllocArgs[] = {Size};
1133   llvm::CallBase *VoidPtr =
1134       CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1135                               CGM.getModule(), OMPRTL___kmpc_alloc_shared),
1136                           AllocArgs, VD->getName());
1137   VoidPtr->addRetAttr(llvm::Attribute::get(
1138       CGM.getLLVMContext(), llvm::Attribute::Alignment, Align.getQuantity()));
1139 
1140   return std::make_pair(VoidPtr, Size);
1141 }
1142 
1143 void CGOpenMPRuntimeGPU::getKmpcFreeShared(
1144     CodeGenFunction &CGF,
1145     const std::pair<llvm::Value *, llvm::Value *> &AddrSizePair) {
1146   // Deallocate the memory for each globalized VLA object
1147   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1148                           CGM.getModule(), OMPRTL___kmpc_free_shared),
1149                       {AddrSizePair.first, AddrSizePair.second});
1150 }
1151 
1152 void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF) {
1153   if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic)
1154     return;
1155 
1156   const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
1157   if (I != FunctionGlobalizedDecls.end()) {
1158     // Deallocate the memory for each globalized VLA object that was
1159     // globalized in the prolog (i.e. emitGenericVarsProlog).
1160     for (const auto &AddrSizePair :
1161          llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) {
1162       CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1163                               CGM.getModule(), OMPRTL___kmpc_free_shared),
1164                           {AddrSizePair.first, AddrSizePair.second});
1165     }
1166     // Deallocate the memory for each globalized value
1167     for (auto &Rec : llvm::reverse(I->getSecond().LocalVarData)) {
1168       const auto *VD = cast<VarDecl>(Rec.first);
1169       I->getSecond().MappedParams->restore(CGF);
1170 
1171       llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal,
1172                                  CGF.getTypeSize(VD->getType())};
1173       CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1174                               CGM.getModule(), OMPRTL___kmpc_free_shared),
1175                           FreeArgs);
1176     }
1177   }
1178 }
1179 
1180 void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF,
1181                                          const OMPExecutableDirective &D,
1182                                          SourceLocation Loc,
1183                                          llvm::Function *OutlinedFn,
1184                                          ArrayRef<llvm::Value *> CapturedVars) {
1185   if (!CGF.HaveInsertPoint())
1186     return;
1187 
1188   bool IsBareKernel = D.getSingleClause<OMPXBareClause>();
1189 
1190   RawAddress ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
1191                                                          /*Name=*/".zero.addr");
1192   CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr);
1193   llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
1194   // We don't emit any thread id function call in bare kernel, but because the
1195   // outlined function has a pointer argument, we emit a nullptr here.
1196   if (IsBareKernel)
1197     OutlinedFnArgs.push_back(llvm::ConstantPointerNull::get(CGM.VoidPtrTy));
1198   else
1199     OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).emitRawPointer(CGF));
1200   OutlinedFnArgs.push_back(ZeroAddr.getPointer());
1201   OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
1202   emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs);
1203 }
1204 
1205 void CGOpenMPRuntimeGPU::emitParallelCall(CodeGenFunction &CGF,
1206                                           SourceLocation Loc,
1207                                           llvm::Function *OutlinedFn,
1208                                           ArrayRef<llvm::Value *> CapturedVars,
1209                                           const Expr *IfCond,
1210                                           llvm::Value *NumThreads) {
1211   if (!CGF.HaveInsertPoint())
1212     return;
1213 
1214   auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond,
1215                         NumThreads](CodeGenFunction &CGF,
1216                                     PrePostActionTy &Action) {
1217     CGBuilderTy &Bld = CGF.Builder;
1218     llvm::Value *NumThreadsVal = NumThreads;
1219     llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn];
1220     llvm::Value *ID = llvm::ConstantPointerNull::get(CGM.Int8PtrTy);
1221     if (WFn)
1222       ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy);
1223     llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(OutlinedFn, CGM.Int8PtrTy);
1224 
1225     // Create a private scope that will globalize the arguments
1226     // passed from the outside of the target region.
1227     // TODO: Is that needed?
1228     CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF);
1229 
1230     Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca(
1231         llvm::ArrayType::get(CGM.VoidPtrTy, CapturedVars.size()),
1232         "captured_vars_addrs");
1233     // There's something to share.
1234     if (!CapturedVars.empty()) {
1235       // Prepare for parallel region. Indicate the outlined function.
1236       ASTContext &Ctx = CGF.getContext();
1237       unsigned Idx = 0;
1238       for (llvm::Value *V : CapturedVars) {
1239         Address Dst = Bld.CreateConstArrayGEP(CapturedVarsAddrs, Idx);
1240         llvm::Value *PtrV;
1241         if (V->getType()->isIntegerTy())
1242           PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy);
1243         else
1244           PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy);
1245         CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false,
1246                               Ctx.getPointerType(Ctx.VoidPtrTy));
1247         ++Idx;
1248       }
1249     }
1250 
1251     llvm::Value *IfCondVal = nullptr;
1252     if (IfCond)
1253       IfCondVal = Bld.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.Int32Ty,
1254                                     /* isSigned */ false);
1255     else
1256       IfCondVal = llvm::ConstantInt::get(CGF.Int32Ty, 1);
1257 
1258     if (!NumThreadsVal)
1259       NumThreadsVal = llvm::ConstantInt::get(CGF.Int32Ty, -1);
1260     else
1261       NumThreadsVal = Bld.CreateZExtOrTrunc(NumThreadsVal, CGF.Int32Ty),
1262 
1263       assert(IfCondVal && "Expected a value");
1264     llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
1265     llvm::Value *Args[] = {
1266         RTLoc,
1267         getThreadID(CGF, Loc),
1268         IfCondVal,
1269         NumThreadsVal,
1270         llvm::ConstantInt::get(CGF.Int32Ty, -1),
1271         FnPtr,
1272         ID,
1273         Bld.CreateBitOrPointerCast(CapturedVarsAddrs.emitRawPointer(CGF),
1274                                    CGF.VoidPtrPtrTy),
1275         llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())};
1276     CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1277                             CGM.getModule(), OMPRTL___kmpc_parallel_51),
1278                         Args);
1279   };
1280 
1281   RegionCodeGenTy RCG(ParallelGen);
1282   RCG(CGF);
1283 }
1284 
1285 void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) {
1286   // Always emit simple barriers!
1287   if (!CGF.HaveInsertPoint())
1288     return;
1289   // Build call __kmpc_barrier_simple_spmd(nullptr, 0);
1290   // This function does not use parameters, so we can emit just default values.
1291   llvm::Value *Args[] = {
1292       llvm::ConstantPointerNull::get(
1293           cast<llvm::PointerType>(getIdentTyPointerTy())),
1294       llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)};
1295   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1296                           CGM.getModule(), OMPRTL___kmpc_barrier_simple_spmd),
1297                       Args);
1298 }
1299 
1300 void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF,
1301                                            SourceLocation Loc,
1302                                            OpenMPDirectiveKind Kind, bool,
1303                                            bool) {
1304   // Always emit simple barriers!
1305   if (!CGF.HaveInsertPoint())
1306     return;
1307   // Build call __kmpc_cancel_barrier(loc, thread_id);
1308   unsigned Flags = getDefaultFlagsForBarriers(Kind);
1309   llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags),
1310                          getThreadID(CGF, Loc)};
1311 
1312   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1313                           CGM.getModule(), OMPRTL___kmpc_barrier),
1314                       Args);
1315 }
1316 
1317 void CGOpenMPRuntimeGPU::emitCriticalRegion(
1318     CodeGenFunction &CGF, StringRef CriticalName,
1319     const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc,
1320     const Expr *Hint) {
1321   llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop");
1322   llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test");
1323   llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync");
1324   llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body");
1325   llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit");
1326 
1327   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1328 
1329   // Get the mask of active threads in the warp.
1330   llvm::Value *Mask = CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1331       CGM.getModule(), OMPRTL___kmpc_warp_active_thread_mask));
1332   // Fetch team-local id of the thread.
1333   llvm::Value *ThreadID = RT.getGPUThreadID(CGF);
1334 
1335   // Get the width of the team.
1336   llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF);
1337 
1338   // Initialize the counter variable for the loop.
1339   QualType Int32Ty =
1340       CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0);
1341   Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter");
1342   LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty);
1343   CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal,
1344                         /*isInit=*/true);
1345 
1346   // Block checks if loop counter exceeds upper bound.
1347   CGF.EmitBlock(LoopBB);
1348   llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
1349   llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth);
1350   CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB);
1351 
1352   // Block tests which single thread should execute region, and which threads
1353   // should go straight to synchronisation point.
1354   CGF.EmitBlock(TestBB);
1355   CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
1356   llvm::Value *CmpThreadToCounter =
1357       CGF.Builder.CreateICmpEQ(ThreadID, CounterVal);
1358   CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB);
1359 
1360   // Block emits the body of the critical region.
1361   CGF.EmitBlock(BodyBB);
1362 
1363   // Output the critical statement.
1364   CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc,
1365                                       Hint);
1366 
1367   // After the body surrounded by the critical region, the single executing
1368   // thread will jump to the synchronisation point.
1369   // Block waits for all threads in current team to finish then increments the
1370   // counter variable and returns to the loop.
1371   CGF.EmitBlock(SyncBB);
1372   // Reconverge active threads in the warp.
1373   (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1374                                 CGM.getModule(), OMPRTL___kmpc_syncwarp),
1375                             Mask);
1376 
1377   llvm::Value *IncCounterVal =
1378       CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1));
1379   CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal);
1380   CGF.EmitBranch(LoopBB);
1381 
1382   // Block that is reached when  all threads in the team complete the region.
1383   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1384 }
1385 
1386 /// Cast value to the specified type.
1387 static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val,
1388                                     QualType ValTy, QualType CastTy,
1389                                     SourceLocation Loc) {
1390   assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() &&
1391          "Cast type must sized.");
1392   assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() &&
1393          "Val type must sized.");
1394   llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy);
1395   if (ValTy == CastTy)
1396     return Val;
1397   if (CGF.getContext().getTypeSizeInChars(ValTy) ==
1398       CGF.getContext().getTypeSizeInChars(CastTy))
1399     return CGF.Builder.CreateBitCast(Val, LLVMCastTy);
1400   if (CastTy->isIntegerType() && ValTy->isIntegerType())
1401     return CGF.Builder.CreateIntCast(Val, LLVMCastTy,
1402                                      CastTy->hasSignedIntegerRepresentation());
1403   Address CastItem = CGF.CreateMemTemp(CastTy);
1404   Address ValCastItem = CastItem.withElementType(Val->getType());
1405   CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy,
1406                         LValueBaseInfo(AlignmentSource::Type),
1407                         TBAAAccessInfo());
1408   return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc,
1409                               LValueBaseInfo(AlignmentSource::Type),
1410                               TBAAAccessInfo());
1411 }
1412 
1413 ///
1414 /// Design of OpenMP reductions on the GPU
1415 ///
1416 /// Consider a typical OpenMP program with one or more reduction
1417 /// clauses:
1418 ///
1419 /// float foo;
1420 /// double bar;
1421 /// #pragma omp target teams distribute parallel for \
1422 ///             reduction(+:foo) reduction(*:bar)
1423 /// for (int i = 0; i < N; i++) {
1424 ///   foo += A[i]; bar *= B[i];
1425 /// }
1426 ///
1427 /// where 'foo' and 'bar' are reduced across all OpenMP threads in
1428 /// all teams.  In our OpenMP implementation on the NVPTX device an
1429 /// OpenMP team is mapped to a CUDA threadblock and OpenMP threads
1430 /// within a team are mapped to CUDA threads within a threadblock.
1431 /// Our goal is to efficiently aggregate values across all OpenMP
1432 /// threads such that:
1433 ///
1434 ///   - the compiler and runtime are logically concise, and
1435 ///   - the reduction is performed efficiently in a hierarchical
1436 ///     manner as follows: within OpenMP threads in the same warp,
1437 ///     across warps in a threadblock, and finally across teams on
1438 ///     the NVPTX device.
1439 ///
1440 /// Introduction to Decoupling
1441 ///
1442 /// We would like to decouple the compiler and the runtime so that the
1443 /// latter is ignorant of the reduction variables (number, data types)
1444 /// and the reduction operators.  This allows a simpler interface
1445 /// and implementation while still attaining good performance.
1446 ///
1447 /// Pseudocode for the aforementioned OpenMP program generated by the
1448 /// compiler is as follows:
1449 ///
1450 /// 1. Create private copies of reduction variables on each OpenMP
1451 ///    thread: 'foo_private', 'bar_private'
1452 /// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned
1453 ///    to it and writes the result in 'foo_private' and 'bar_private'
1454 ///    respectively.
1455 /// 3. Call the OpenMP runtime on the GPU to reduce within a team
1456 ///    and store the result on the team master:
1457 ///
1458 ///     __kmpc_nvptx_parallel_reduce_nowait_v2(...,
1459 ///        reduceData, shuffleReduceFn, interWarpCpyFn)
1460 ///
1461 ///     where:
1462 ///       struct ReduceData {
1463 ///         double *foo;
1464 ///         double *bar;
1465 ///       } reduceData
1466 ///       reduceData.foo = &foo_private
1467 ///       reduceData.bar = &bar_private
1468 ///
1469 ///     'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two
1470 ///     auxiliary functions generated by the compiler that operate on
1471 ///     variables of type 'ReduceData'.  They aid the runtime perform
1472 ///     algorithmic steps in a data agnostic manner.
1473 ///
1474 ///     'shuffleReduceFn' is a pointer to a function that reduces data
1475 ///     of type 'ReduceData' across two OpenMP threads (lanes) in the
1476 ///     same warp.  It takes the following arguments as input:
1477 ///
1478 ///     a. variable of type 'ReduceData' on the calling lane,
1479 ///     b. its lane_id,
1480 ///     c. an offset relative to the current lane_id to generate a
1481 ///        remote_lane_id.  The remote lane contains the second
1482 ///        variable of type 'ReduceData' that is to be reduced.
1483 ///     d. an algorithm version parameter determining which reduction
1484 ///        algorithm to use.
1485 ///
1486 ///     'shuffleReduceFn' retrieves data from the remote lane using
1487 ///     efficient GPU shuffle intrinsics and reduces, using the
1488 ///     algorithm specified by the 4th parameter, the two operands
1489 ///     element-wise.  The result is written to the first operand.
1490 ///
1491 ///     Different reduction algorithms are implemented in different
1492 ///     runtime functions, all calling 'shuffleReduceFn' to perform
1493 ///     the essential reduction step.  Therefore, based on the 4th
1494 ///     parameter, this function behaves slightly differently to
1495 ///     cooperate with the runtime to ensure correctness under
1496 ///     different circumstances.
1497 ///
1498 ///     'InterWarpCpyFn' is a pointer to a function that transfers
1499 ///     reduced variables across warps.  It tunnels, through CUDA
1500 ///     shared memory, the thread-private data of type 'ReduceData'
1501 ///     from lane 0 of each warp to a lane in the first warp.
1502 /// 4. Call the OpenMP runtime on the GPU to reduce across teams.
1503 ///    The last team writes the global reduced value to memory.
1504 ///
1505 ///     ret = __kmpc_nvptx_teams_reduce_nowait(...,
1506 ///             reduceData, shuffleReduceFn, interWarpCpyFn,
1507 ///             scratchpadCopyFn, loadAndReduceFn)
1508 ///
1509 ///     'scratchpadCopyFn' is a helper that stores reduced
1510 ///     data from the team master to a scratchpad array in
1511 ///     global memory.
1512 ///
1513 ///     'loadAndReduceFn' is a helper that loads data from
1514 ///     the scratchpad array and reduces it with the input
1515 ///     operand.
1516 ///
1517 ///     These compiler generated functions hide address
1518 ///     calculation and alignment information from the runtime.
1519 /// 5. if ret == 1:
1520 ///     The team master of the last team stores the reduced
1521 ///     result to the globals in memory.
1522 ///     foo += reduceData.foo; bar *= reduceData.bar
1523 ///
1524 ///
1525 /// Warp Reduction Algorithms
1526 ///
1527 /// On the warp level, we have three algorithms implemented in the
1528 /// OpenMP runtime depending on the number of active lanes:
1529 ///
1530 /// Full Warp Reduction
1531 ///
1532 /// The reduce algorithm within a warp where all lanes are active
1533 /// is implemented in the runtime as follows:
1534 ///
1535 /// full_warp_reduce(void *reduce_data,
1536 ///                  kmp_ShuffleReductFctPtr ShuffleReduceFn) {
1537 ///   for (int offset = WARPSIZE/2; offset > 0; offset /= 2)
1538 ///     ShuffleReduceFn(reduce_data, 0, offset, 0);
1539 /// }
1540 ///
1541 /// The algorithm completes in log(2, WARPSIZE) steps.
1542 ///
1543 /// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is
1544 /// not used therefore we save instructions by not retrieving lane_id
1545 /// from the corresponding special registers.  The 4th parameter, which
1546 /// represents the version of the algorithm being used, is set to 0 to
1547 /// signify full warp reduction.
1548 ///
1549 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
1550 ///
1551 /// #reduce_elem refers to an element in the local lane's data structure
1552 /// #remote_elem is retrieved from a remote lane
1553 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
1554 /// reduce_elem = reduce_elem REDUCE_OP remote_elem;
1555 ///
1556 /// Contiguous Partial Warp Reduction
1557 ///
1558 /// This reduce algorithm is used within a warp where only the first
1559 /// 'n' (n <= WARPSIZE) lanes are active.  It is typically used when the
1560 /// number of OpenMP threads in a parallel region is not a multiple of
1561 /// WARPSIZE.  The algorithm is implemented in the runtime as follows:
1562 ///
1563 /// void
1564 /// contiguous_partial_reduce(void *reduce_data,
1565 ///                           kmp_ShuffleReductFctPtr ShuffleReduceFn,
1566 ///                           int size, int lane_id) {
1567 ///   int curr_size;
1568 ///   int offset;
1569 ///   curr_size = size;
1570 ///   mask = curr_size/2;
1571 ///   while (offset>0) {
1572 ///     ShuffleReduceFn(reduce_data, lane_id, offset, 1);
1573 ///     curr_size = (curr_size+1)/2;
1574 ///     offset = curr_size/2;
1575 ///   }
1576 /// }
1577 ///
1578 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
1579 ///
1580 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
1581 /// if (lane_id < offset)
1582 ///     reduce_elem = reduce_elem REDUCE_OP remote_elem
1583 /// else
1584 ///     reduce_elem = remote_elem
1585 ///
1586 /// This algorithm assumes that the data to be reduced are located in a
1587 /// contiguous subset of lanes starting from the first.  When there is
1588 /// an odd number of active lanes, the data in the last lane is not
1589 /// aggregated with any other lane's dat but is instead copied over.
1590 ///
1591 /// Dispersed Partial Warp Reduction
1592 ///
1593 /// This algorithm is used within a warp when any discontiguous subset of
1594 /// lanes are active.  It is used to implement the reduction operation
1595 /// across lanes in an OpenMP simd region or in a nested parallel region.
1596 ///
1597 /// void
1598 /// dispersed_partial_reduce(void *reduce_data,
1599 ///                          kmp_ShuffleReductFctPtr ShuffleReduceFn) {
1600 ///   int size, remote_id;
1601 ///   int logical_lane_id = number_of_active_lanes_before_me() * 2;
1602 ///   do {
1603 ///       remote_id = next_active_lane_id_right_after_me();
1604 ///       # the above function returns 0 of no active lane
1605 ///       # is present right after the current lane.
1606 ///       size = number_of_active_lanes_in_this_warp();
1607 ///       logical_lane_id /= 2;
1608 ///       ShuffleReduceFn(reduce_data, logical_lane_id,
1609 ///                       remote_id-1-threadIdx.x, 2);
1610 ///   } while (logical_lane_id % 2 == 0 && size > 1);
1611 /// }
1612 ///
1613 /// There is no assumption made about the initial state of the reduction.
1614 /// Any number of lanes (>=1) could be active at any position.  The reduction
1615 /// result is returned in the first active lane.
1616 ///
1617 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
1618 ///
1619 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
1620 /// if (lane_id % 2 == 0 && offset > 0)
1621 ///     reduce_elem = reduce_elem REDUCE_OP remote_elem
1622 /// else
1623 ///     reduce_elem = remote_elem
1624 ///
1625 ///
1626 /// Intra-Team Reduction
1627 ///
1628 /// This function, as implemented in the runtime call
1629 /// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP
1630 /// threads in a team.  It first reduces within a warp using the
1631 /// aforementioned algorithms.  We then proceed to gather all such
1632 /// reduced values at the first warp.
1633 ///
1634 /// The runtime makes use of the function 'InterWarpCpyFn', which copies
1635 /// data from each of the "warp master" (zeroth lane of each warp, where
1636 /// warp-reduced data is held) to the zeroth warp.  This step reduces (in
1637 /// a mathematical sense) the problem of reduction across warp masters in
1638 /// a block to the problem of warp reduction.
1639 ///
1640 ///
1641 /// Inter-Team Reduction
1642 ///
1643 /// Once a team has reduced its data to a single value, it is stored in
1644 /// a global scratchpad array.  Since each team has a distinct slot, this
1645 /// can be done without locking.
1646 ///
1647 /// The last team to write to the scratchpad array proceeds to reduce the
1648 /// scratchpad array.  One or more workers in the last team use the helper
1649 /// 'loadAndReduceDataFn' to load and reduce values from the array, i.e.,
1650 /// the k'th worker reduces every k'th element.
1651 ///
1652 /// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to
1653 /// reduce across workers and compute a globally reduced value.
1654 ///
1655 void CGOpenMPRuntimeGPU::emitReduction(
1656     CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates,
1657     ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs,
1658     ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) {
1659   if (!CGF.HaveInsertPoint())
1660     return;
1661 
1662   bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind);
1663   bool DistributeReduction = isOpenMPDistributeDirective(Options.ReductionKind);
1664   bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind);
1665 
1666   ASTContext &C = CGM.getContext();
1667 
1668   if (Options.SimpleReduction) {
1669     assert(!TeamsReduction && !ParallelReduction &&
1670            "Invalid reduction selection in emitReduction.");
1671     (void)ParallelReduction;
1672     CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs,
1673                                    ReductionOps, Options);
1674     return;
1675   }
1676 
1677   llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap;
1678   llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size());
1679   int Cnt = 0;
1680   for (const Expr *DRE : Privates) {
1681     PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl();
1682     ++Cnt;
1683   }
1684   const RecordDecl *ReductionRec = ::buildRecordForGlobalizedVars(
1685       CGM.getContext(), PrivatesReductions, std::nullopt, VarFieldMap, 1);
1686 
1687   if (TeamsReduction)
1688     TeamsReductions.push_back(ReductionRec);
1689 
1690   // Source location for the ident struct
1691   llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
1692 
1693   using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy;
1694   InsertPointTy AllocaIP(CGF.AllocaInsertPt->getParent(),
1695                          CGF.AllocaInsertPt->getIterator());
1696   InsertPointTy CodeGenIP(CGF.Builder.GetInsertBlock(),
1697                           CGF.Builder.GetInsertPoint());
1698   llvm::OpenMPIRBuilder::LocationDescription OmpLoc(
1699       CodeGenIP, CGF.SourceLocToDebugLoc(Loc));
1700   llvm::SmallVector<llvm::OpenMPIRBuilder::ReductionInfo> ReductionInfos;
1701 
1702   CodeGenFunction::OMPPrivateScope Scope(CGF);
1703   unsigned Idx = 0;
1704   for (const Expr *Private : Privates) {
1705     llvm::Type *ElementType;
1706     llvm::Value *Variable;
1707     llvm::Value *PrivateVariable;
1708     llvm::OpenMPIRBuilder::ReductionGenAtomicCBTy AtomicReductionGen = nullptr;
1709     ElementType = CGF.ConvertTypeForMem(Private->getType());
1710     const auto *RHSVar =
1711         cast<VarDecl>(cast<DeclRefExpr>(RHSExprs[Idx])->getDecl());
1712     PrivateVariable = CGF.GetAddrOfLocalVar(RHSVar).emitRawPointer(CGF);
1713     const auto *LHSVar =
1714         cast<VarDecl>(cast<DeclRefExpr>(LHSExprs[Idx])->getDecl());
1715     Variable = CGF.GetAddrOfLocalVar(LHSVar).emitRawPointer(CGF);
1716     llvm::OpenMPIRBuilder::EvalKind EvalKind;
1717     switch (CGF.getEvaluationKind(Private->getType())) {
1718     case TEK_Scalar:
1719       EvalKind = llvm::OpenMPIRBuilder::EvalKind::Scalar;
1720       break;
1721     case TEK_Complex:
1722       EvalKind = llvm::OpenMPIRBuilder::EvalKind::Complex;
1723       break;
1724     case TEK_Aggregate:
1725       EvalKind = llvm::OpenMPIRBuilder::EvalKind::Aggregate;
1726       break;
1727     }
1728     auto ReductionGen = [&](InsertPointTy CodeGenIP, unsigned I,
1729                             llvm::Value **LHSPtr, llvm::Value **RHSPtr,
1730                             llvm::Function *NewFunc) {
1731       CGF.Builder.restoreIP(CodeGenIP);
1732       auto *CurFn = CGF.CurFn;
1733       CGF.CurFn = NewFunc;
1734 
1735       *LHSPtr = CGF.GetAddrOfLocalVar(
1736                        cast<VarDecl>(cast<DeclRefExpr>(LHSExprs[I])->getDecl()))
1737                     .emitRawPointer(CGF);
1738       *RHSPtr = CGF.GetAddrOfLocalVar(
1739                        cast<VarDecl>(cast<DeclRefExpr>(RHSExprs[I])->getDecl()))
1740                     .emitRawPointer(CGF);
1741 
1742       emitSingleReductionCombiner(CGF, ReductionOps[I], Privates[I],
1743                                   cast<DeclRefExpr>(LHSExprs[I]),
1744                                   cast<DeclRefExpr>(RHSExprs[I]));
1745 
1746       CGF.CurFn = CurFn;
1747 
1748       return InsertPointTy(CGF.Builder.GetInsertBlock(),
1749                            CGF.Builder.GetInsertPoint());
1750     };
1751     ReductionInfos.emplace_back(llvm::OpenMPIRBuilder::ReductionInfo(
1752         ElementType, Variable, PrivateVariable, EvalKind,
1753         /*ReductionGen=*/nullptr, ReductionGen, AtomicReductionGen));
1754     Idx++;
1755   }
1756 
1757   CGF.Builder.restoreIP(OMPBuilder.createReductionsGPU(
1758       OmpLoc, AllocaIP, CodeGenIP, ReductionInfos, false, TeamsReduction,
1759       DistributeReduction, llvm::OpenMPIRBuilder::ReductionGenCBKind::Clang,
1760       CGF.getTarget().getGridValue(), C.getLangOpts().OpenMPCUDAReductionBufNum,
1761       RTLoc));
1762   return;
1763 }
1764 
1765 const VarDecl *
1766 CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD,
1767                                        const VarDecl *NativeParam) const {
1768   if (!NativeParam->getType()->isReferenceType())
1769     return NativeParam;
1770   QualType ArgType = NativeParam->getType();
1771   QualifierCollector QC;
1772   const Type *NonQualTy = QC.strip(ArgType);
1773   QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
1774   if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) {
1775     if (Attr->getCaptureKind() == OMPC_map) {
1776       PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy,
1777                                                         LangAS::opencl_global);
1778     }
1779   }
1780   ArgType = CGM.getContext().getPointerType(PointeeTy);
1781   QC.addRestrict();
1782   enum { NVPTX_local_addr = 5 };
1783   QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr));
1784   ArgType = QC.apply(CGM.getContext(), ArgType);
1785   if (isa<ImplicitParamDecl>(NativeParam))
1786     return ImplicitParamDecl::Create(
1787         CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(),
1788         NativeParam->getIdentifier(), ArgType, ImplicitParamKind::Other);
1789   return ParmVarDecl::Create(
1790       CGM.getContext(),
1791       const_cast<DeclContext *>(NativeParam->getDeclContext()),
1792       NativeParam->getBeginLoc(), NativeParam->getLocation(),
1793       NativeParam->getIdentifier(), ArgType,
1794       /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr);
1795 }
1796 
1797 Address
1798 CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF,
1799                                           const VarDecl *NativeParam,
1800                                           const VarDecl *TargetParam) const {
1801   assert(NativeParam != TargetParam &&
1802          NativeParam->getType()->isReferenceType() &&
1803          "Native arg must not be the same as target arg.");
1804   Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam);
1805   QualType NativeParamType = NativeParam->getType();
1806   QualifierCollector QC;
1807   const Type *NonQualTy = QC.strip(NativeParamType);
1808   QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
1809   unsigned NativePointeeAddrSpace =
1810       CGF.getTypes().getTargetAddressSpace(NativePointeeTy);
1811   QualType TargetTy = TargetParam->getType();
1812   llvm::Value *TargetAddr = CGF.EmitLoadOfScalar(LocalAddr, /*Volatile=*/false,
1813                                                  TargetTy, SourceLocation());
1814   // Cast to native address space.
1815   TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
1816       TargetAddr,
1817       llvm::PointerType::get(CGF.getLLVMContext(), NativePointeeAddrSpace));
1818   Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType);
1819   CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false,
1820                         NativeParamType);
1821   return NativeParamAddr;
1822 }
1823 
1824 void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall(
1825     CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn,
1826     ArrayRef<llvm::Value *> Args) const {
1827   SmallVector<llvm::Value *, 4> TargetArgs;
1828   TargetArgs.reserve(Args.size());
1829   auto *FnType = OutlinedFn.getFunctionType();
1830   for (unsigned I = 0, E = Args.size(); I < E; ++I) {
1831     if (FnType->isVarArg() && FnType->getNumParams() <= I) {
1832       TargetArgs.append(std::next(Args.begin(), I), Args.end());
1833       break;
1834     }
1835     llvm::Type *TargetType = FnType->getParamType(I);
1836     llvm::Value *NativeArg = Args[I];
1837     if (!TargetType->isPointerTy()) {
1838       TargetArgs.emplace_back(NativeArg);
1839       continue;
1840     }
1841     TargetArgs.emplace_back(
1842         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(NativeArg, TargetType));
1843   }
1844   CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs);
1845 }
1846 
1847 /// Emit function which wraps the outline parallel region
1848 /// and controls the arguments which are passed to this function.
1849 /// The wrapper ensures that the outlined function is called
1850 /// with the correct arguments when data is shared.
1851 llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper(
1852     llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) {
1853   ASTContext &Ctx = CGM.getContext();
1854   const auto &CS = *D.getCapturedStmt(OMPD_parallel);
1855 
1856   // Create a function that takes as argument the source thread.
1857   FunctionArgList WrapperArgs;
1858   QualType Int16QTy =
1859       Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false);
1860   QualType Int32QTy =
1861       Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false);
1862   ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
1863                                      /*Id=*/nullptr, Int16QTy,
1864                                      ImplicitParamKind::Other);
1865   ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
1866                                /*Id=*/nullptr, Int32QTy,
1867                                ImplicitParamKind::Other);
1868   WrapperArgs.emplace_back(&ParallelLevelArg);
1869   WrapperArgs.emplace_back(&WrapperArg);
1870 
1871   const CGFunctionInfo &CGFI =
1872       CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs);
1873 
1874   auto *Fn = llvm::Function::Create(
1875       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
1876       Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule());
1877 
1878   // Ensure we do not inline the function. This is trivially true for the ones
1879   // passed to __kmpc_fork_call but the ones calles in serialized regions
1880   // could be inlined. This is not a perfect but it is closer to the invariant
1881   // we want, namely, every data environment starts with a new function.
1882   // TODO: We should pass the if condition to the runtime function and do the
1883   //       handling there. Much cleaner code.
1884   Fn->addFnAttr(llvm::Attribute::NoInline);
1885 
1886   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
1887   Fn->setLinkage(llvm::GlobalValue::InternalLinkage);
1888   Fn->setDoesNotRecurse();
1889 
1890   CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
1891   CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs,
1892                     D.getBeginLoc(), D.getBeginLoc());
1893 
1894   const auto *RD = CS.getCapturedRecordDecl();
1895   auto CurField = RD->field_begin();
1896 
1897   Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
1898                                                       /*Name=*/".zero.addr");
1899   CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr);
1900   // Get the array of arguments.
1901   SmallVector<llvm::Value *, 8> Args;
1902 
1903   Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).emitRawPointer(CGF));
1904   Args.emplace_back(ZeroAddr.emitRawPointer(CGF));
1905 
1906   CGBuilderTy &Bld = CGF.Builder;
1907   auto CI = CS.capture_begin();
1908 
1909   // Use global memory for data sharing.
1910   // Handle passing of global args to workers.
1911   RawAddress GlobalArgs =
1912       CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args");
1913   llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer();
1914   llvm::Value *DataSharingArgs[] = {GlobalArgsPtr};
1915   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1916                           CGM.getModule(), OMPRTL___kmpc_get_shared_variables),
1917                       DataSharingArgs);
1918 
1919   // Retrieve the shared variables from the list of references returned
1920   // by the runtime. Pass the variables to the outlined function.
1921   Address SharedArgListAddress = Address::invalid();
1922   if (CS.capture_size() > 0 ||
1923       isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
1924     SharedArgListAddress = CGF.EmitLoadOfPointer(
1925         GlobalArgs, CGF.getContext()
1926                         .getPointerType(CGF.getContext().VoidPtrTy)
1927                         .castAs<PointerType>());
1928   }
1929   unsigned Idx = 0;
1930   if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
1931     Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
1932     Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
1933         Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy);
1934     llvm::Value *LB = CGF.EmitLoadOfScalar(
1935         TypedAddress,
1936         /*Volatile=*/false,
1937         CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
1938         cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc());
1939     Args.emplace_back(LB);
1940     ++Idx;
1941     Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
1942     TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
1943         Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy);
1944     llvm::Value *UB = CGF.EmitLoadOfScalar(
1945         TypedAddress,
1946         /*Volatile=*/false,
1947         CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
1948         cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc());
1949     Args.emplace_back(UB);
1950     ++Idx;
1951   }
1952   if (CS.capture_size() > 0) {
1953     ASTContext &CGFContext = CGF.getContext();
1954     for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) {
1955       QualType ElemTy = CurField->getType();
1956       Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx);
1957       Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
1958           Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)),
1959           CGF.ConvertTypeForMem(ElemTy));
1960       llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress,
1961                                               /*Volatile=*/false,
1962                                               CGFContext.getPointerType(ElemTy),
1963                                               CI->getLocation());
1964       if (CI->capturesVariableByCopy() &&
1965           !CI->getCapturedVar()->getType()->isAnyPointerType()) {
1966         Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(),
1967                               CI->getLocation());
1968       }
1969       Args.emplace_back(Arg);
1970     }
1971   }
1972 
1973   emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args);
1974   CGF.FinishFunction();
1975   return Fn;
1976 }
1977 
1978 void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF,
1979                                               const Decl *D) {
1980   if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic)
1981     return;
1982 
1983   assert(D && "Expected function or captured|block decl.");
1984   assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 &&
1985          "Function is registered already.");
1986   assert((!TeamAndReductions.first || TeamAndReductions.first == D) &&
1987          "Team is set but not processed.");
1988   const Stmt *Body = nullptr;
1989   bool NeedToDelayGlobalization = false;
1990   if (const auto *FD = dyn_cast<FunctionDecl>(D)) {
1991     Body = FD->getBody();
1992   } else if (const auto *BD = dyn_cast<BlockDecl>(D)) {
1993     Body = BD->getBody();
1994   } else if (const auto *CD = dyn_cast<CapturedDecl>(D)) {
1995     Body = CD->getBody();
1996     NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP;
1997     if (NeedToDelayGlobalization &&
1998         getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
1999       return;
2000   }
2001   if (!Body)
2002     return;
2003   CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second);
2004   VarChecker.Visit(Body);
2005   const RecordDecl *GlobalizedVarsRecord =
2006       VarChecker.getGlobalizedRecord(IsInTTDRegion);
2007   TeamAndReductions.first = nullptr;
2008   TeamAndReductions.second.clear();
2009   ArrayRef<const ValueDecl *> EscapedVariableLengthDecls =
2010       VarChecker.getEscapedVariableLengthDecls();
2011   ArrayRef<const ValueDecl *> DelayedVariableLengthDecls =
2012       VarChecker.getDelayedVariableLengthDecls();
2013   if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty() &&
2014       DelayedVariableLengthDecls.empty())
2015     return;
2016   auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
2017   I->getSecond().MappedParams =
2018       std::make_unique<CodeGenFunction::OMPMapVars>();
2019   I->getSecond().EscapedParameters.insert(
2020       VarChecker.getEscapedParameters().begin(),
2021       VarChecker.getEscapedParameters().end());
2022   I->getSecond().EscapedVariableLengthDecls.append(
2023       EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end());
2024   I->getSecond().DelayedVariableLengthDecls.append(
2025       DelayedVariableLengthDecls.begin(), DelayedVariableLengthDecls.end());
2026   DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
2027   for (const ValueDecl *VD : VarChecker.getEscapedDecls()) {
2028     assert(VD->isCanonicalDecl() && "Expected canonical declaration");
2029     Data.insert(std::make_pair(VD, MappedVarData()));
2030   }
2031   if (!NeedToDelayGlobalization) {
2032     emitGenericVarsProlog(CGF, D->getBeginLoc());
2033     struct GlobalizationScope final : EHScopeStack::Cleanup {
2034       GlobalizationScope() = default;
2035 
2036       void Emit(CodeGenFunction &CGF, Flags flags) override {
2037         static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
2038             .emitGenericVarsEpilog(CGF);
2039       }
2040     };
2041     CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup);
2042   }
2043 }
2044 
2045 Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF,
2046                                                         const VarDecl *VD) {
2047   if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) {
2048     const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
2049     auto AS = LangAS::Default;
2050     switch (A->getAllocatorType()) {
2051       // Use the default allocator here as by default local vars are
2052       // threadlocal.
2053     case OMPAllocateDeclAttr::OMPNullMemAlloc:
2054     case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
2055     case OMPAllocateDeclAttr::OMPThreadMemAlloc:
2056     case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
2057     case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
2058       // Follow the user decision - use default allocation.
2059       return Address::invalid();
2060     case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
2061       // TODO: implement aupport for user-defined allocators.
2062       return Address::invalid();
2063     case OMPAllocateDeclAttr::OMPConstMemAlloc:
2064       AS = LangAS::cuda_constant;
2065       break;
2066     case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
2067       AS = LangAS::cuda_shared;
2068       break;
2069     case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
2070     case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
2071       break;
2072     }
2073     llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
2074     auto *GV = new llvm::GlobalVariable(
2075         CGM.getModule(), VarTy, /*isConstant=*/false,
2076         llvm::GlobalValue::InternalLinkage, llvm::PoisonValue::get(VarTy),
2077         VD->getName(),
2078         /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
2079         CGM.getContext().getTargetAddressSpace(AS));
2080     CharUnits Align = CGM.getContext().getDeclAlign(VD);
2081     GV->setAlignment(Align.getAsAlign());
2082     return Address(
2083         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2084             GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace(
2085                     VD->getType().getAddressSpace()))),
2086         VarTy, Align);
2087   }
2088 
2089   if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic)
2090     return Address::invalid();
2091 
2092   VD = VD->getCanonicalDecl();
2093   auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
2094   if (I == FunctionGlobalizedDecls.end())
2095     return Address::invalid();
2096   auto VDI = I->getSecond().LocalVarData.find(VD);
2097   if (VDI != I->getSecond().LocalVarData.end())
2098     return VDI->second.PrivateAddr;
2099   if (VD->hasAttrs()) {
2100     for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()),
2101          E(VD->attr_end());
2102          IT != E; ++IT) {
2103       auto VDI = I->getSecond().LocalVarData.find(
2104           cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl())
2105               ->getCanonicalDecl());
2106       if (VDI != I->getSecond().LocalVarData.end())
2107         return VDI->second.PrivateAddr;
2108     }
2109   }
2110 
2111   return Address::invalid();
2112 }
2113 
2114 void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) {
2115   FunctionGlobalizedDecls.erase(CGF.CurFn);
2116   CGOpenMPRuntime::functionFinished(CGF);
2117 }
2118 
2119 void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk(
2120     CodeGenFunction &CGF, const OMPLoopDirective &S,
2121     OpenMPDistScheduleClauseKind &ScheduleKind,
2122     llvm::Value *&Chunk) const {
2123   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
2124   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
2125     ScheduleKind = OMPC_DIST_SCHEDULE_static;
2126     Chunk = CGF.EmitScalarConversion(
2127         RT.getGPUNumThreads(CGF),
2128         CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
2129         S.getIterationVariable()->getType(), S.getBeginLoc());
2130     return;
2131   }
2132   CGOpenMPRuntime::getDefaultDistScheduleAndChunk(
2133       CGF, S, ScheduleKind, Chunk);
2134 }
2135 
2136 void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk(
2137     CodeGenFunction &CGF, const OMPLoopDirective &S,
2138     OpenMPScheduleClauseKind &ScheduleKind,
2139     const Expr *&ChunkExpr) const {
2140   ScheduleKind = OMPC_SCHEDULE_static;
2141   // Chunk size is 1 in this case.
2142   llvm::APInt ChunkSize(32, 1);
2143   ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize,
2144       CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
2145       SourceLocation());
2146 }
2147 
2148 void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas(
2149     CodeGenFunction &CGF, const OMPExecutableDirective &D) const {
2150   assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&
2151          " Expected target-based directive.");
2152   const CapturedStmt *CS = D.getCapturedStmt(OMPD_target);
2153   for (const CapturedStmt::Capture &C : CS->captures()) {
2154     // Capture variables captured by reference in lambdas for target-based
2155     // directives.
2156     if (!C.capturesVariable())
2157       continue;
2158     const VarDecl *VD = C.getCapturedVar();
2159     const auto *RD = VD->getType()
2160                          .getCanonicalType()
2161                          .getNonReferenceType()
2162                          ->getAsCXXRecordDecl();
2163     if (!RD || !RD->isLambda())
2164       continue;
2165     Address VDAddr = CGF.GetAddrOfLocalVar(VD);
2166     LValue VDLVal;
2167     if (VD->getType().getCanonicalType()->isReferenceType())
2168       VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType());
2169     else
2170       VDLVal = CGF.MakeAddrLValue(
2171           VDAddr, VD->getType().getCanonicalType().getNonReferenceType());
2172     llvm::DenseMap<const ValueDecl *, FieldDecl *> Captures;
2173     FieldDecl *ThisCapture = nullptr;
2174     RD->getCaptureFields(Captures, ThisCapture);
2175     if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) {
2176       LValue ThisLVal =
2177           CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture);
2178       llvm::Value *CXXThis = CGF.LoadCXXThis();
2179       CGF.EmitStoreOfScalar(CXXThis, ThisLVal);
2180     }
2181     for (const LambdaCapture &LC : RD->captures()) {
2182       if (LC.getCaptureKind() != LCK_ByRef)
2183         continue;
2184       const ValueDecl *VD = LC.getCapturedVar();
2185       // FIXME: For now VD is always a VarDecl because OpenMP does not support
2186       //  capturing structured bindings in lambdas yet.
2187       if (!CS->capturesVariable(cast<VarDecl>(VD)))
2188         continue;
2189       auto It = Captures.find(VD);
2190       assert(It != Captures.end() && "Found lambda capture without field.");
2191       LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second);
2192       Address VDAddr = CGF.GetAddrOfLocalVar(cast<VarDecl>(VD));
2193       if (VD->getType().getCanonicalType()->isReferenceType())
2194         VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr,
2195                                                VD->getType().getCanonicalType())
2196                      .getAddress();
2197       CGF.EmitStoreOfScalar(VDAddr.emitRawPointer(CGF), VarLVal);
2198     }
2199   }
2200 }
2201 
2202 bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD,
2203                                                             LangAS &AS) {
2204   if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())
2205     return false;
2206   const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
2207   switch(A->getAllocatorType()) {
2208   case OMPAllocateDeclAttr::OMPNullMemAlloc:
2209   case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
2210   // Not supported, fallback to the default mem space.
2211   case OMPAllocateDeclAttr::OMPThreadMemAlloc:
2212   case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
2213   case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
2214   case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
2215   case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
2216     AS = LangAS::Default;
2217     return true;
2218   case OMPAllocateDeclAttr::OMPConstMemAlloc:
2219     AS = LangAS::cuda_constant;
2220     return true;
2221   case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
2222     AS = LangAS::cuda_shared;
2223     return true;
2224   case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
2225     llvm_unreachable("Expected predefined allocator for the variables with the "
2226                      "static storage.");
2227   }
2228   return false;
2229 }
2230 
2231 // Get current OffloadArch and ignore any unknown values
2232 static OffloadArch getOffloadArch(CodeGenModule &CGM) {
2233   if (!CGM.getTarget().hasFeature("ptx"))
2234     return OffloadArch::UNKNOWN;
2235   for (const auto &Feature : CGM.getTarget().getTargetOpts().FeatureMap) {
2236     if (Feature.getValue()) {
2237       OffloadArch Arch = StringToOffloadArch(Feature.getKey());
2238       if (Arch != OffloadArch::UNKNOWN)
2239         return Arch;
2240     }
2241   }
2242   return OffloadArch::UNKNOWN;
2243 }
2244 
2245 /// Check to see if target architecture supports unified addressing which is
2246 /// a restriction for OpenMP requires clause "unified_shared_memory".
2247 void CGOpenMPRuntimeGPU::processRequiresDirective(const OMPRequiresDecl *D) {
2248   for (const OMPClause *Clause : D->clauselists()) {
2249     if (Clause->getClauseKind() == OMPC_unified_shared_memory) {
2250       OffloadArch Arch = getOffloadArch(CGM);
2251       switch (Arch) {
2252       case OffloadArch::SM_20:
2253       case OffloadArch::SM_21:
2254       case OffloadArch::SM_30:
2255       case OffloadArch::SM_32_:
2256       case OffloadArch::SM_35:
2257       case OffloadArch::SM_37:
2258       case OffloadArch::SM_50:
2259       case OffloadArch::SM_52:
2260       case OffloadArch::SM_53: {
2261         SmallString<256> Buffer;
2262         llvm::raw_svector_ostream Out(Buffer);
2263         Out << "Target architecture " << OffloadArchToString(Arch)
2264             << " does not support unified addressing";
2265         CGM.Error(Clause->getBeginLoc(), Out.str());
2266         return;
2267       }
2268       case OffloadArch::SM_60:
2269       case OffloadArch::SM_61:
2270       case OffloadArch::SM_62:
2271       case OffloadArch::SM_70:
2272       case OffloadArch::SM_72:
2273       case OffloadArch::SM_75:
2274       case OffloadArch::SM_80:
2275       case OffloadArch::SM_86:
2276       case OffloadArch::SM_87:
2277       case OffloadArch::SM_89:
2278       case OffloadArch::SM_90:
2279       case OffloadArch::SM_90a:
2280       case OffloadArch::GFX600:
2281       case OffloadArch::GFX601:
2282       case OffloadArch::GFX602:
2283       case OffloadArch::GFX700:
2284       case OffloadArch::GFX701:
2285       case OffloadArch::GFX702:
2286       case OffloadArch::GFX703:
2287       case OffloadArch::GFX704:
2288       case OffloadArch::GFX705:
2289       case OffloadArch::GFX801:
2290       case OffloadArch::GFX802:
2291       case OffloadArch::GFX803:
2292       case OffloadArch::GFX805:
2293       case OffloadArch::GFX810:
2294       case OffloadArch::GFX9_GENERIC:
2295       case OffloadArch::GFX900:
2296       case OffloadArch::GFX902:
2297       case OffloadArch::GFX904:
2298       case OffloadArch::GFX906:
2299       case OffloadArch::GFX908:
2300       case OffloadArch::GFX909:
2301       case OffloadArch::GFX90a:
2302       case OffloadArch::GFX90c:
2303       case OffloadArch::GFX940:
2304       case OffloadArch::GFX941:
2305       case OffloadArch::GFX942:
2306       case OffloadArch::GFX10_1_GENERIC:
2307       case OffloadArch::GFX1010:
2308       case OffloadArch::GFX1011:
2309       case OffloadArch::GFX1012:
2310       case OffloadArch::GFX1013:
2311       case OffloadArch::GFX10_3_GENERIC:
2312       case OffloadArch::GFX1030:
2313       case OffloadArch::GFX1031:
2314       case OffloadArch::GFX1032:
2315       case OffloadArch::GFX1033:
2316       case OffloadArch::GFX1034:
2317       case OffloadArch::GFX1035:
2318       case OffloadArch::GFX1036:
2319       case OffloadArch::GFX11_GENERIC:
2320       case OffloadArch::GFX1100:
2321       case OffloadArch::GFX1101:
2322       case OffloadArch::GFX1102:
2323       case OffloadArch::GFX1103:
2324       case OffloadArch::GFX1150:
2325       case OffloadArch::GFX1151:
2326       case OffloadArch::GFX1152:
2327       case OffloadArch::GFX12_GENERIC:
2328       case OffloadArch::GFX1200:
2329       case OffloadArch::GFX1201:
2330       case OffloadArch::AMDGCNSPIRV:
2331       case OffloadArch::Generic:
2332       case OffloadArch::UNUSED:
2333       case OffloadArch::UNKNOWN:
2334         break;
2335       case OffloadArch::LAST:
2336         llvm_unreachable("Unexpected GPU arch.");
2337       }
2338     }
2339   }
2340   CGOpenMPRuntime::processRequiresDirective(D);
2341 }
2342 
2343 llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) {
2344   CGBuilderTy &Bld = CGF.Builder;
2345   llvm::Module *M = &CGF.CGM.getModule();
2346   const char *LocSize = "__kmpc_get_hardware_num_threads_in_block";
2347   llvm::Function *F = M->getFunction(LocSize);
2348   if (!F) {
2349     F = llvm::Function::Create(
2350         llvm::FunctionType::get(CGF.Int32Ty, std::nullopt, false),
2351         llvm::GlobalVariable::ExternalLinkage, LocSize, &CGF.CGM.getModule());
2352   }
2353   return Bld.CreateCall(F, std::nullopt, "nvptx_num_threads");
2354 }
2355 
2356 llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) {
2357   ArrayRef<llvm::Value *> Args{};
2358   return CGF.EmitRuntimeCall(
2359       OMPBuilder.getOrCreateRuntimeFunction(
2360           CGM.getModule(), OMPRTL___kmpc_get_hardware_thread_id_in_block),
2361       Args);
2362 }
2363