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/StmtOpenMP.h"
19 #include "clang/AST/StmtVisitor.h"
20 #include "clang/Basic/Cuda.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/Frontend/OpenMP/OMPGridValues.h"
23 #include "llvm/Support/MathExtras.h"
24 
25 using namespace clang;
26 using namespace CodeGen;
27 using namespace llvm::omp;
28 
29 namespace {
30 /// Pre(post)-action for different OpenMP constructs specialized for NVPTX.
31 class NVPTXActionTy final : public PrePostActionTy {
32   llvm::FunctionCallee EnterCallee = nullptr;
33   ArrayRef<llvm::Value *> EnterArgs;
34   llvm::FunctionCallee ExitCallee = nullptr;
35   ArrayRef<llvm::Value *> ExitArgs;
36   bool Conditional = false;
37   llvm::BasicBlock *ContBlock = nullptr;
38 
39 public:
40   NVPTXActionTy(llvm::FunctionCallee EnterCallee,
41                 ArrayRef<llvm::Value *> EnterArgs,
42                 llvm::FunctionCallee ExitCallee,
43                 ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false)
44       : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee),
45         ExitArgs(ExitArgs), Conditional(Conditional) {}
46   void Enter(CodeGenFunction &CGF) override {
47     llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs);
48     if (Conditional) {
49       llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes);
50       auto *ThenBlock = CGF.createBasicBlock("omp_if.then");
51       ContBlock = CGF.createBasicBlock("omp_if.end");
52       // Generate the branch (If-stmt)
53       CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock);
54       CGF.EmitBlock(ThenBlock);
55     }
56   }
57   void Done(CodeGenFunction &CGF) {
58     // Emit the rest of blocks/branches
59     CGF.EmitBranch(ContBlock);
60     CGF.EmitBlock(ContBlock, true);
61   }
62   void Exit(CodeGenFunction &CGF) override {
63     CGF.EmitRuntimeCall(ExitCallee, ExitArgs);
64   }
65 };
66 
67 /// A class to track the execution mode when codegening directives within
68 /// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry
69 /// to the target region and used by containing directives such as 'parallel'
70 /// to emit optimized code.
71 class ExecutionRuntimeModesRAII {
72 private:
73   CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode =
74       CGOpenMPRuntimeGPU::EM_Unknown;
75   CGOpenMPRuntimeGPU::ExecutionMode &ExecMode;
76   bool SavedRuntimeMode = false;
77   bool *RuntimeMode = nullptr;
78 
79 public:
80   /// Constructor for Non-SPMD mode.
81   ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode)
82       : ExecMode(ExecMode) {
83     SavedExecMode = ExecMode;
84     ExecMode = CGOpenMPRuntimeGPU::EM_NonSPMD;
85   }
86   /// Constructor for SPMD mode.
87   ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode,
88                             bool &RuntimeMode, bool FullRuntimeMode)
89       : ExecMode(ExecMode), RuntimeMode(&RuntimeMode) {
90     SavedExecMode = ExecMode;
91     SavedRuntimeMode = RuntimeMode;
92     ExecMode = CGOpenMPRuntimeGPU::EM_SPMD;
93     RuntimeMode = FullRuntimeMode;
94   }
95   ~ExecutionRuntimeModesRAII() {
96     ExecMode = SavedExecMode;
97     if (RuntimeMode)
98       *RuntimeMode = SavedRuntimeMode;
99   }
100 };
101 
102 /// GPU Configuration:  This information can be derived from cuda registers,
103 /// however, providing compile time constants helps generate more efficient
104 /// code.  For all practical purposes this is fine because the configuration
105 /// is the same for all known NVPTX architectures.
106 enum MachineConfiguration : unsigned {
107   /// See "llvm/Frontend/OpenMP/OMPGridValues.h" for various related target
108   /// specific Grid Values like GV_Warp_Size, GV_Slot_Size
109 
110   /// Global memory alignment for performance.
111   GlobalMemoryAlignment = 128,
112 
113   /// Maximal size of the shared memory buffer.
114   SharedMemorySize = 128,
115 };
116 
117 static const ValueDecl *getPrivateItem(const Expr *RefExpr) {
118   RefExpr = RefExpr->IgnoreParens();
119   if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) {
120     const Expr *Base = ASE->getBase()->IgnoreParenImpCasts();
121     while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
122       Base = TempASE->getBase()->IgnoreParenImpCasts();
123     RefExpr = Base;
124   } else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) {
125     const Expr *Base = OASE->getBase()->IgnoreParenImpCasts();
126     while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base))
127       Base = TempOASE->getBase()->IgnoreParenImpCasts();
128     while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
129       Base = TempASE->getBase()->IgnoreParenImpCasts();
130     RefExpr = Base;
131   }
132   RefExpr = RefExpr->IgnoreParenImpCasts();
133   if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr))
134     return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl());
135   const auto *ME = cast<MemberExpr>(RefExpr);
136   return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl());
137 }
138 
139 
140 static RecordDecl *buildRecordForGlobalizedVars(
141     ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls,
142     ArrayRef<const ValueDecl *> EscapedDeclsForTeams,
143     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
144         &MappedDeclsFields, int BufSize) {
145   using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>;
146   if (EscapedDecls.empty() && EscapedDeclsForTeams.empty())
147     return nullptr;
148   SmallVector<VarsDataTy, 4> GlobalizedVars;
149   for (const ValueDecl *D : EscapedDecls)
150     GlobalizedVars.emplace_back(
151         CharUnits::fromQuantity(std::max(
152             C.getDeclAlign(D).getQuantity(),
153             static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))),
154         D);
155   for (const ValueDecl *D : EscapedDeclsForTeams)
156     GlobalizedVars.emplace_back(C.getDeclAlign(D), D);
157   llvm::stable_sort(GlobalizedVars, [](VarsDataTy L, VarsDataTy R) {
158     return L.first > R.first;
159   });
160 
161   // Build struct _globalized_locals_ty {
162   //         /*  globalized vars  */[WarSize] align (max(decl_align,
163   //         GlobalMemoryAlignment))
164   //         /*  globalized vars  */ for EscapedDeclsForTeams
165   //       };
166   RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty");
167   GlobalizedRD->startDefinition();
168   llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped(
169       EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end());
170   for (const auto &Pair : GlobalizedVars) {
171     const ValueDecl *VD = Pair.second;
172     QualType Type = VD->getType();
173     if (Type->isLValueReferenceType())
174       Type = C.getPointerType(Type.getNonReferenceType());
175     else
176       Type = Type.getNonReferenceType();
177     SourceLocation Loc = VD->getLocation();
178     FieldDecl *Field;
179     if (SingleEscaped.count(VD)) {
180       Field = FieldDecl::Create(
181           C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
182           C.getTrivialTypeSourceInfo(Type, SourceLocation()),
183           /*BW=*/nullptr, /*Mutable=*/false,
184           /*InitStyle=*/ICIS_NoInit);
185       Field->setAccess(AS_public);
186       if (VD->hasAttrs()) {
187         for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
188              E(VD->getAttrs().end());
189              I != E; ++I)
190           Field->addAttr(*I);
191       }
192     } else {
193       llvm::APInt ArraySize(32, BufSize);
194       Type = C.getConstantArrayType(Type, ArraySize, nullptr, ArrayType::Normal,
195                                     0);
196       Field = FieldDecl::Create(
197           C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
198           C.getTrivialTypeSourceInfo(Type, SourceLocation()),
199           /*BW=*/nullptr, /*Mutable=*/false,
200           /*InitStyle=*/ICIS_NoInit);
201       Field->setAccess(AS_public);
202       llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(),
203                                      static_cast<CharUnits::QuantityType>(
204                                          GlobalMemoryAlignment)));
205       Field->addAttr(AlignedAttr::CreateImplicit(
206           C, /*IsAlignmentExpr=*/true,
207           IntegerLiteral::Create(C, Align,
208                                  C.getIntTypeForBitwidth(32, /*Signed=*/0),
209                                  SourceLocation()),
210           {}, AttributeCommonInfo::AS_GNU, AlignedAttr::GNU_aligned));
211     }
212     GlobalizedRD->addDecl(Field);
213     MappedDeclsFields.try_emplace(VD, Field);
214   }
215   GlobalizedRD->completeDefinition();
216   return GlobalizedRD;
217 }
218 
219 /// Get the list of variables that can escape their declaration context.
220 class CheckVarsEscapingDeclContext final
221     : public ConstStmtVisitor<CheckVarsEscapingDeclContext> {
222   CodeGenFunction &CGF;
223   llvm::SetVector<const ValueDecl *> EscapedDecls;
224   llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls;
225   llvm::SmallPtrSet<const Decl *, 4> EscapedParameters;
226   RecordDecl *GlobalizedRD = nullptr;
227   llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
228   bool AllEscaped = false;
229   bool IsForCombinedParallelRegion = false;
230 
231   void markAsEscaped(const ValueDecl *VD) {
232     // Do not globalize declare target variables.
233     if (!isa<VarDecl>(VD) ||
234         OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD))
235       return;
236     VD = cast<ValueDecl>(VD->getCanonicalDecl());
237     // Use user-specified allocation.
238     if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>())
239       return;
240     // Variables captured by value must be globalized.
241     if (auto *CSI = CGF.CapturedStmtInfo) {
242       if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) {
243         // Check if need to capture the variable that was already captured by
244         // value in the outer region.
245         if (!IsForCombinedParallelRegion) {
246           if (!FD->hasAttrs())
247             return;
248           const auto *Attr = FD->getAttr<OMPCaptureKindAttr>();
249           if (!Attr)
250             return;
251           if (((Attr->getCaptureKind() != OMPC_map) &&
252                !isOpenMPPrivate(Attr->getCaptureKind())) ||
253               ((Attr->getCaptureKind() == OMPC_map) &&
254                !FD->getType()->isAnyPointerType()))
255             return;
256         }
257         if (!FD->getType()->isReferenceType()) {
258           assert(!VD->getType()->isVariablyModifiedType() &&
259                  "Parameter captured by value with variably modified type");
260           EscapedParameters.insert(VD);
261         } else if (!IsForCombinedParallelRegion) {
262           return;
263         }
264       }
265     }
266     if ((!CGF.CapturedStmtInfo ||
267          (IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) &&
268         VD->getType()->isReferenceType())
269       // Do not globalize variables with reference type.
270       return;
271     if (VD->getType()->isVariablyModifiedType())
272       EscapedVariableLengthDecls.insert(VD);
273     else
274       EscapedDecls.insert(VD);
275   }
276 
277   void VisitValueDecl(const ValueDecl *VD) {
278     if (VD->getType()->isLValueReferenceType())
279       markAsEscaped(VD);
280     if (const auto *VarD = dyn_cast<VarDecl>(VD)) {
281       if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) {
282         const bool SavedAllEscaped = AllEscaped;
283         AllEscaped = VD->getType()->isLValueReferenceType();
284         Visit(VarD->getInit());
285         AllEscaped = SavedAllEscaped;
286       }
287     }
288   }
289   void VisitOpenMPCapturedStmt(const CapturedStmt *S,
290                                ArrayRef<OMPClause *> Clauses,
291                                bool IsCombinedParallelRegion) {
292     if (!S)
293       return;
294     for (const CapturedStmt::Capture &C : S->captures()) {
295       if (C.capturesVariable() && !C.capturesVariableByCopy()) {
296         const ValueDecl *VD = C.getCapturedVar();
297         bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion;
298         if (IsCombinedParallelRegion) {
299           // Check if the variable is privatized in the combined construct and
300           // those private copies must be shared in the inner parallel
301           // directive.
302           IsForCombinedParallelRegion = false;
303           for (const OMPClause *C : Clauses) {
304             if (!isOpenMPPrivate(C->getClauseKind()) ||
305                 C->getClauseKind() == OMPC_reduction ||
306                 C->getClauseKind() == OMPC_linear ||
307                 C->getClauseKind() == OMPC_private)
308               continue;
309             ArrayRef<const Expr *> Vars;
310             if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C))
311               Vars = PC->getVarRefs();
312             else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C))
313               Vars = PC->getVarRefs();
314             else
315               llvm_unreachable("Unexpected clause.");
316             for (const auto *E : Vars) {
317               const Decl *D =
318                   cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl();
319               if (D == VD->getCanonicalDecl()) {
320                 IsForCombinedParallelRegion = true;
321                 break;
322               }
323             }
324             if (IsForCombinedParallelRegion)
325               break;
326           }
327         }
328         markAsEscaped(VD);
329         if (isa<OMPCapturedExprDecl>(VD))
330           VisitValueDecl(VD);
331         IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion;
332       }
333     }
334   }
335 
336   void buildRecordForGlobalizedVars(bool IsInTTDRegion) {
337     assert(!GlobalizedRD &&
338            "Record for globalized variables is built already.");
339     ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams;
340     unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size;
341     if (IsInTTDRegion)
342       EscapedDeclsForTeams = EscapedDecls.getArrayRef();
343     else
344       EscapedDeclsForParallel = EscapedDecls.getArrayRef();
345     GlobalizedRD = ::buildRecordForGlobalizedVars(
346         CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams,
347         MappedDeclsFields, WarpSize);
348   }
349 
350 public:
351   CheckVarsEscapingDeclContext(CodeGenFunction &CGF,
352                                ArrayRef<const ValueDecl *> TeamsReductions)
353       : CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) {
354   }
355   virtual ~CheckVarsEscapingDeclContext() = default;
356   void VisitDeclStmt(const DeclStmt *S) {
357     if (!S)
358       return;
359     for (const Decl *D : S->decls())
360       if (const auto *VD = dyn_cast_or_null<ValueDecl>(D))
361         VisitValueDecl(VD);
362   }
363   void VisitOMPExecutableDirective(const OMPExecutableDirective *D) {
364     if (!D)
365       return;
366     if (!D->hasAssociatedStmt())
367       return;
368     if (const auto *S =
369             dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) {
370       // Do not analyze directives that do not actually require capturing,
371       // like `omp for` or `omp simd` directives.
372       llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
373       getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind());
374       if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) {
375         VisitStmt(S->getCapturedStmt());
376         return;
377       }
378       VisitOpenMPCapturedStmt(
379           S, D->clauses(),
380           CaptureRegions.back() == OMPD_parallel &&
381               isOpenMPDistributeDirective(D->getDirectiveKind()));
382     }
383   }
384   void VisitCapturedStmt(const CapturedStmt *S) {
385     if (!S)
386       return;
387     for (const CapturedStmt::Capture &C : S->captures()) {
388       if (C.capturesVariable() && !C.capturesVariableByCopy()) {
389         const ValueDecl *VD = C.getCapturedVar();
390         markAsEscaped(VD);
391         if (isa<OMPCapturedExprDecl>(VD))
392           VisitValueDecl(VD);
393       }
394     }
395   }
396   void VisitLambdaExpr(const LambdaExpr *E) {
397     if (!E)
398       return;
399     for (const LambdaCapture &C : E->captures()) {
400       if (C.capturesVariable()) {
401         if (C.getCaptureKind() == LCK_ByRef) {
402           const ValueDecl *VD = C.getCapturedVar();
403           markAsEscaped(VD);
404           if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD))
405             VisitValueDecl(VD);
406         }
407       }
408     }
409   }
410   void VisitBlockExpr(const BlockExpr *E) {
411     if (!E)
412       return;
413     for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) {
414       if (C.isByRef()) {
415         const VarDecl *VD = C.getVariable();
416         markAsEscaped(VD);
417         if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture())
418           VisitValueDecl(VD);
419       }
420     }
421   }
422   void VisitCallExpr(const CallExpr *E) {
423     if (!E)
424       return;
425     for (const Expr *Arg : E->arguments()) {
426       if (!Arg)
427         continue;
428       if (Arg->isLValue()) {
429         const bool SavedAllEscaped = AllEscaped;
430         AllEscaped = true;
431         Visit(Arg);
432         AllEscaped = SavedAllEscaped;
433       } else {
434         Visit(Arg);
435       }
436     }
437     Visit(E->getCallee());
438   }
439   void VisitDeclRefExpr(const DeclRefExpr *E) {
440     if (!E)
441       return;
442     const ValueDecl *VD = E->getDecl();
443     if (AllEscaped)
444       markAsEscaped(VD);
445     if (isa<OMPCapturedExprDecl>(VD))
446       VisitValueDecl(VD);
447     else if (const auto *VarD = dyn_cast<VarDecl>(VD))
448       if (VarD->isInitCapture())
449         VisitValueDecl(VD);
450   }
451   void VisitUnaryOperator(const UnaryOperator *E) {
452     if (!E)
453       return;
454     if (E->getOpcode() == UO_AddrOf) {
455       const bool SavedAllEscaped = AllEscaped;
456       AllEscaped = true;
457       Visit(E->getSubExpr());
458       AllEscaped = SavedAllEscaped;
459     } else {
460       Visit(E->getSubExpr());
461     }
462   }
463   void VisitImplicitCastExpr(const ImplicitCastExpr *E) {
464     if (!E)
465       return;
466     if (E->getCastKind() == CK_ArrayToPointerDecay) {
467       const bool SavedAllEscaped = AllEscaped;
468       AllEscaped = true;
469       Visit(E->getSubExpr());
470       AllEscaped = SavedAllEscaped;
471     } else {
472       Visit(E->getSubExpr());
473     }
474   }
475   void VisitExpr(const Expr *E) {
476     if (!E)
477       return;
478     bool SavedAllEscaped = AllEscaped;
479     if (!E->isLValue())
480       AllEscaped = false;
481     for (const Stmt *Child : E->children())
482       if (Child)
483         Visit(Child);
484     AllEscaped = SavedAllEscaped;
485   }
486   void VisitStmt(const Stmt *S) {
487     if (!S)
488       return;
489     for (const Stmt *Child : S->children())
490       if (Child)
491         Visit(Child);
492   }
493 
494   /// Returns the record that handles all the escaped local variables and used
495   /// instead of their original storage.
496   const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) {
497     if (!GlobalizedRD)
498       buildRecordForGlobalizedVars(IsInTTDRegion);
499     return GlobalizedRD;
500   }
501 
502   /// Returns the field in the globalized record for the escaped variable.
503   const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const {
504     assert(GlobalizedRD &&
505            "Record for globalized variables must be generated already.");
506     auto I = MappedDeclsFields.find(VD);
507     if (I == MappedDeclsFields.end())
508       return nullptr;
509     return I->getSecond();
510   }
511 
512   /// Returns the list of the escaped local variables/parameters.
513   ArrayRef<const ValueDecl *> getEscapedDecls() const {
514     return EscapedDecls.getArrayRef();
515   }
516 
517   /// Checks if the escaped local variable is actually a parameter passed by
518   /// value.
519   const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const {
520     return EscapedParameters;
521   }
522 
523   /// Returns the list of the escaped variables with the variably modified
524   /// types.
525   ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const {
526     return EscapedVariableLengthDecls.getArrayRef();
527   }
528 };
529 } // anonymous namespace
530 
531 /// Get the id of the warp in the block.
532 /// We assume that the warp size is 32, which is always the case
533 /// on the NVPTX device, to generate more efficient code.
534 static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) {
535   CGBuilderTy &Bld = CGF.Builder;
536   unsigned LaneIDBits =
537       llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size);
538   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
539   return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id");
540 }
541 
542 /// Get the id of the current lane in the Warp.
543 /// We assume that the warp size is 32, which is always the case
544 /// on the NVPTX device, to generate more efficient code.
545 static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) {
546   CGBuilderTy &Bld = CGF.Builder;
547   unsigned LaneIDBits =
548       llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size);
549   unsigned LaneIDMask = ~0u >> (32u - LaneIDBits);
550   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
551   return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask),
552                        "nvptx_lane_id");
553 }
554 
555 CGOpenMPRuntimeGPU::ExecutionMode
556 CGOpenMPRuntimeGPU::getExecutionMode() const {
557   return CurrentExecutionMode;
558 }
559 
560 static CGOpenMPRuntimeGPU::DataSharingMode
561 getDataSharingMode(CodeGenModule &CGM) {
562   return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeGPU::CUDA
563                                           : CGOpenMPRuntimeGPU::Generic;
564 }
565 
566 /// Check for inner (nested) SPMD construct, if any
567 static bool hasNestedSPMDDirective(ASTContext &Ctx,
568                                    const OMPExecutableDirective &D) {
569   const auto *CS = D.getInnermostCapturedStmt();
570   const auto *Body =
571       CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
572   const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
573 
574   if (const auto *NestedDir =
575           dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
576     OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
577     switch (D.getDirectiveKind()) {
578     case OMPD_target:
579       if (isOpenMPParallelDirective(DKind))
580         return true;
581       if (DKind == OMPD_teams) {
582         Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
583             /*IgnoreCaptured=*/true);
584         if (!Body)
585           return false;
586         ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
587         if (const auto *NND =
588                 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
589           DKind = NND->getDirectiveKind();
590           if (isOpenMPParallelDirective(DKind))
591             return true;
592         }
593       }
594       return false;
595     case OMPD_target_teams:
596       return isOpenMPParallelDirective(DKind);
597     case OMPD_target_simd:
598     case OMPD_target_parallel:
599     case OMPD_target_parallel_for:
600     case OMPD_target_parallel_for_simd:
601     case OMPD_target_teams_distribute:
602     case OMPD_target_teams_distribute_simd:
603     case OMPD_target_teams_distribute_parallel_for:
604     case OMPD_target_teams_distribute_parallel_for_simd:
605     case OMPD_parallel:
606     case OMPD_for:
607     case OMPD_parallel_for:
608     case OMPD_parallel_master:
609     case OMPD_parallel_sections:
610     case OMPD_for_simd:
611     case OMPD_parallel_for_simd:
612     case OMPD_cancel:
613     case OMPD_cancellation_point:
614     case OMPD_ordered:
615     case OMPD_threadprivate:
616     case OMPD_allocate:
617     case OMPD_task:
618     case OMPD_simd:
619     case OMPD_sections:
620     case OMPD_section:
621     case OMPD_single:
622     case OMPD_master:
623     case OMPD_critical:
624     case OMPD_taskyield:
625     case OMPD_barrier:
626     case OMPD_taskwait:
627     case OMPD_taskgroup:
628     case OMPD_atomic:
629     case OMPD_flush:
630     case OMPD_depobj:
631     case OMPD_scan:
632     case OMPD_teams:
633     case OMPD_target_data:
634     case OMPD_target_exit_data:
635     case OMPD_target_enter_data:
636     case OMPD_distribute:
637     case OMPD_distribute_simd:
638     case OMPD_distribute_parallel_for:
639     case OMPD_distribute_parallel_for_simd:
640     case OMPD_teams_distribute:
641     case OMPD_teams_distribute_simd:
642     case OMPD_teams_distribute_parallel_for:
643     case OMPD_teams_distribute_parallel_for_simd:
644     case OMPD_target_update:
645     case OMPD_declare_simd:
646     case OMPD_declare_variant:
647     case OMPD_begin_declare_variant:
648     case OMPD_end_declare_variant:
649     case OMPD_declare_target:
650     case OMPD_end_declare_target:
651     case OMPD_declare_reduction:
652     case OMPD_declare_mapper:
653     case OMPD_taskloop:
654     case OMPD_taskloop_simd:
655     case OMPD_master_taskloop:
656     case OMPD_master_taskloop_simd:
657     case OMPD_parallel_master_taskloop:
658     case OMPD_parallel_master_taskloop_simd:
659     case OMPD_requires:
660     case OMPD_unknown:
661     default:
662       llvm_unreachable("Unexpected directive.");
663     }
664   }
665 
666   return false;
667 }
668 
669 static bool supportsSPMDExecutionMode(ASTContext &Ctx,
670                                       const OMPExecutableDirective &D) {
671   OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
672   switch (DirectiveKind) {
673   case OMPD_target:
674   case OMPD_target_teams:
675     return hasNestedSPMDDirective(Ctx, D);
676   case OMPD_target_parallel:
677   case OMPD_target_parallel_for:
678   case OMPD_target_parallel_for_simd:
679   case OMPD_target_teams_distribute_parallel_for:
680   case OMPD_target_teams_distribute_parallel_for_simd:
681   case OMPD_target_simd:
682   case OMPD_target_teams_distribute_simd:
683     return true;
684   case OMPD_target_teams_distribute:
685     return false;
686   case OMPD_parallel:
687   case OMPD_for:
688   case OMPD_parallel_for:
689   case OMPD_parallel_master:
690   case OMPD_parallel_sections:
691   case OMPD_for_simd:
692   case OMPD_parallel_for_simd:
693   case OMPD_cancel:
694   case OMPD_cancellation_point:
695   case OMPD_ordered:
696   case OMPD_threadprivate:
697   case OMPD_allocate:
698   case OMPD_task:
699   case OMPD_simd:
700   case OMPD_sections:
701   case OMPD_section:
702   case OMPD_single:
703   case OMPD_master:
704   case OMPD_critical:
705   case OMPD_taskyield:
706   case OMPD_barrier:
707   case OMPD_taskwait:
708   case OMPD_taskgroup:
709   case OMPD_atomic:
710   case OMPD_flush:
711   case OMPD_depobj:
712   case OMPD_scan:
713   case OMPD_teams:
714   case OMPD_target_data:
715   case OMPD_target_exit_data:
716   case OMPD_target_enter_data:
717   case OMPD_distribute:
718   case OMPD_distribute_simd:
719   case OMPD_distribute_parallel_for:
720   case OMPD_distribute_parallel_for_simd:
721   case OMPD_teams_distribute:
722   case OMPD_teams_distribute_simd:
723   case OMPD_teams_distribute_parallel_for:
724   case OMPD_teams_distribute_parallel_for_simd:
725   case OMPD_target_update:
726   case OMPD_declare_simd:
727   case OMPD_declare_variant:
728   case OMPD_begin_declare_variant:
729   case OMPD_end_declare_variant:
730   case OMPD_declare_target:
731   case OMPD_end_declare_target:
732   case OMPD_declare_reduction:
733   case OMPD_declare_mapper:
734   case OMPD_taskloop:
735   case OMPD_taskloop_simd:
736   case OMPD_master_taskloop:
737   case OMPD_master_taskloop_simd:
738   case OMPD_parallel_master_taskloop:
739   case OMPD_parallel_master_taskloop_simd:
740   case OMPD_requires:
741   case OMPD_unknown:
742   default:
743     break;
744   }
745   llvm_unreachable(
746       "Unknown programming model for OpenMP directive on NVPTX target.");
747 }
748 
749 /// Check if the directive is loops based and has schedule clause at all or has
750 /// static scheduling.
751 static bool hasStaticScheduling(const OMPExecutableDirective &D) {
752   assert(isOpenMPWorksharingDirective(D.getDirectiveKind()) &&
753          isOpenMPLoopDirective(D.getDirectiveKind()) &&
754          "Expected loop-based directive.");
755   return !D.hasClausesOfKind<OMPOrderedClause>() &&
756          (!D.hasClausesOfKind<OMPScheduleClause>() ||
757           llvm::any_of(D.getClausesOfKind<OMPScheduleClause>(),
758                        [](const OMPScheduleClause *C) {
759                          return C->getScheduleKind() == OMPC_SCHEDULE_static;
760                        }));
761 }
762 
763 /// Check for inner (nested) lightweight runtime construct, if any
764 static bool hasNestedLightweightDirective(ASTContext &Ctx,
765                                           const OMPExecutableDirective &D) {
766   assert(supportsSPMDExecutionMode(Ctx, D) && "Expected SPMD mode directive.");
767   const auto *CS = D.getInnermostCapturedStmt();
768   const auto *Body =
769       CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
770   const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
771 
772   if (const auto *NestedDir =
773           dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
774     OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
775     switch (D.getDirectiveKind()) {
776     case OMPD_target:
777       if (isOpenMPParallelDirective(DKind) &&
778           isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) &&
779           hasStaticScheduling(*NestedDir))
780         return true;
781       if (DKind == OMPD_teams_distribute_simd || DKind == OMPD_simd)
782         return true;
783       if (DKind == OMPD_parallel) {
784         Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
785             /*IgnoreCaptured=*/true);
786         if (!Body)
787           return false;
788         ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
789         if (const auto *NND =
790                 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
791           DKind = NND->getDirectiveKind();
792           if (isOpenMPWorksharingDirective(DKind) &&
793               isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
794             return true;
795         }
796       } else if (DKind == OMPD_teams) {
797         Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
798             /*IgnoreCaptured=*/true);
799         if (!Body)
800           return false;
801         ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
802         if (const auto *NND =
803                 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
804           DKind = NND->getDirectiveKind();
805           if (isOpenMPParallelDirective(DKind) &&
806               isOpenMPWorksharingDirective(DKind) &&
807               isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
808             return true;
809           if (DKind == OMPD_parallel) {
810             Body = NND->getInnermostCapturedStmt()->IgnoreContainers(
811                 /*IgnoreCaptured=*/true);
812             if (!Body)
813               return false;
814             ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
815             if (const auto *NND =
816                     dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
817               DKind = NND->getDirectiveKind();
818               if (isOpenMPWorksharingDirective(DKind) &&
819                   isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
820                 return true;
821             }
822           }
823         }
824       }
825       return false;
826     case OMPD_target_teams:
827       if (isOpenMPParallelDirective(DKind) &&
828           isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) &&
829           hasStaticScheduling(*NestedDir))
830         return true;
831       if (DKind == OMPD_distribute_simd || DKind == OMPD_simd)
832         return true;
833       if (DKind == OMPD_parallel) {
834         Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
835             /*IgnoreCaptured=*/true);
836         if (!Body)
837           return false;
838         ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
839         if (const auto *NND =
840                 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
841           DKind = NND->getDirectiveKind();
842           if (isOpenMPWorksharingDirective(DKind) &&
843               isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND))
844             return true;
845         }
846       }
847       return false;
848     case OMPD_target_parallel:
849       if (DKind == OMPD_simd)
850         return true;
851       return isOpenMPWorksharingDirective(DKind) &&
852              isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NestedDir);
853     case OMPD_target_teams_distribute:
854     case OMPD_target_simd:
855     case OMPD_target_parallel_for:
856     case OMPD_target_parallel_for_simd:
857     case OMPD_target_teams_distribute_simd:
858     case OMPD_target_teams_distribute_parallel_for:
859     case OMPD_target_teams_distribute_parallel_for_simd:
860     case OMPD_parallel:
861     case OMPD_for:
862     case OMPD_parallel_for:
863     case OMPD_parallel_master:
864     case OMPD_parallel_sections:
865     case OMPD_for_simd:
866     case OMPD_parallel_for_simd:
867     case OMPD_cancel:
868     case OMPD_cancellation_point:
869     case OMPD_ordered:
870     case OMPD_threadprivate:
871     case OMPD_allocate:
872     case OMPD_task:
873     case OMPD_simd:
874     case OMPD_sections:
875     case OMPD_section:
876     case OMPD_single:
877     case OMPD_master:
878     case OMPD_critical:
879     case OMPD_taskyield:
880     case OMPD_barrier:
881     case OMPD_taskwait:
882     case OMPD_taskgroup:
883     case OMPD_atomic:
884     case OMPD_flush:
885     case OMPD_depobj:
886     case OMPD_scan:
887     case OMPD_teams:
888     case OMPD_target_data:
889     case OMPD_target_exit_data:
890     case OMPD_target_enter_data:
891     case OMPD_distribute:
892     case OMPD_distribute_simd:
893     case OMPD_distribute_parallel_for:
894     case OMPD_distribute_parallel_for_simd:
895     case OMPD_teams_distribute:
896     case OMPD_teams_distribute_simd:
897     case OMPD_teams_distribute_parallel_for:
898     case OMPD_teams_distribute_parallel_for_simd:
899     case OMPD_target_update:
900     case OMPD_declare_simd:
901     case OMPD_declare_variant:
902     case OMPD_begin_declare_variant:
903     case OMPD_end_declare_variant:
904     case OMPD_declare_target:
905     case OMPD_end_declare_target:
906     case OMPD_declare_reduction:
907     case OMPD_declare_mapper:
908     case OMPD_taskloop:
909     case OMPD_taskloop_simd:
910     case OMPD_master_taskloop:
911     case OMPD_master_taskloop_simd:
912     case OMPD_parallel_master_taskloop:
913     case OMPD_parallel_master_taskloop_simd:
914     case OMPD_requires:
915     case OMPD_unknown:
916     default:
917       llvm_unreachable("Unexpected directive.");
918     }
919   }
920 
921   return false;
922 }
923 
924 /// Checks if the construct supports lightweight runtime. It must be SPMD
925 /// construct + inner loop-based construct with static scheduling.
926 static bool supportsLightweightRuntime(ASTContext &Ctx,
927                                        const OMPExecutableDirective &D) {
928   if (!supportsSPMDExecutionMode(Ctx, D))
929     return false;
930   OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
931   switch (DirectiveKind) {
932   case OMPD_target:
933   case OMPD_target_teams:
934   case OMPD_target_parallel:
935     return hasNestedLightweightDirective(Ctx, D);
936   case OMPD_target_parallel_for:
937   case OMPD_target_parallel_for_simd:
938   case OMPD_target_teams_distribute_parallel_for:
939   case OMPD_target_teams_distribute_parallel_for_simd:
940     // (Last|First)-privates must be shared in parallel region.
941     return hasStaticScheduling(D);
942   case OMPD_target_simd:
943   case OMPD_target_teams_distribute_simd:
944     return true;
945   case OMPD_target_teams_distribute:
946     return false;
947   case OMPD_parallel:
948   case OMPD_for:
949   case OMPD_parallel_for:
950   case OMPD_parallel_master:
951   case OMPD_parallel_sections:
952   case OMPD_for_simd:
953   case OMPD_parallel_for_simd:
954   case OMPD_cancel:
955   case OMPD_cancellation_point:
956   case OMPD_ordered:
957   case OMPD_threadprivate:
958   case OMPD_allocate:
959   case OMPD_task:
960   case OMPD_simd:
961   case OMPD_sections:
962   case OMPD_section:
963   case OMPD_single:
964   case OMPD_master:
965   case OMPD_critical:
966   case OMPD_taskyield:
967   case OMPD_barrier:
968   case OMPD_taskwait:
969   case OMPD_taskgroup:
970   case OMPD_atomic:
971   case OMPD_flush:
972   case OMPD_depobj:
973   case OMPD_scan:
974   case OMPD_teams:
975   case OMPD_target_data:
976   case OMPD_target_exit_data:
977   case OMPD_target_enter_data:
978   case OMPD_distribute:
979   case OMPD_distribute_simd:
980   case OMPD_distribute_parallel_for:
981   case OMPD_distribute_parallel_for_simd:
982   case OMPD_teams_distribute:
983   case OMPD_teams_distribute_simd:
984   case OMPD_teams_distribute_parallel_for:
985   case OMPD_teams_distribute_parallel_for_simd:
986   case OMPD_target_update:
987   case OMPD_declare_simd:
988   case OMPD_declare_variant:
989   case OMPD_begin_declare_variant:
990   case OMPD_end_declare_variant:
991   case OMPD_declare_target:
992   case OMPD_end_declare_target:
993   case OMPD_declare_reduction:
994   case OMPD_declare_mapper:
995   case OMPD_taskloop:
996   case OMPD_taskloop_simd:
997   case OMPD_master_taskloop:
998   case OMPD_master_taskloop_simd:
999   case OMPD_parallel_master_taskloop:
1000   case OMPD_parallel_master_taskloop_simd:
1001   case OMPD_requires:
1002   case OMPD_unknown:
1003   default:
1004     break;
1005   }
1006   llvm_unreachable(
1007       "Unknown programming model for OpenMP directive on NVPTX target.");
1008 }
1009 
1010 void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D,
1011                                              StringRef ParentName,
1012                                              llvm::Function *&OutlinedFn,
1013                                              llvm::Constant *&OutlinedFnID,
1014                                              bool IsOffloadEntry,
1015                                              const RegionCodeGenTy &CodeGen) {
1016   ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode);
1017   EntryFunctionState EST;
1018   WrapperFunctionsMap.clear();
1019 
1020   // Emit target region as a standalone region.
1021   class NVPTXPrePostActionTy : public PrePostActionTy {
1022     CGOpenMPRuntimeGPU::EntryFunctionState &EST;
1023 
1024   public:
1025     NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST)
1026         : EST(EST) {}
1027     void Enter(CodeGenFunction &CGF) override {
1028       auto &RT =
1029           static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1030       RT.emitKernelInit(CGF, EST, /* IsSPMD */ false);
1031       // Skip target region initialization.
1032       RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
1033     }
1034     void Exit(CodeGenFunction &CGF) override {
1035       auto &RT =
1036           static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1037       RT.clearLocThreadIdInsertPt(CGF);
1038       RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ false);
1039     }
1040   } Action(EST);
1041   CodeGen.setAction(Action);
1042   IsInTTDRegion = true;
1043   emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
1044                                    IsOffloadEntry, CodeGen);
1045   IsInTTDRegion = false;
1046 }
1047 
1048 void CGOpenMPRuntimeGPU::emitKernelInit(CodeGenFunction &CGF,
1049                                         EntryFunctionState &EST, bool IsSPMD) {
1050   CGBuilderTy &Bld = CGF.Builder;
1051   Bld.restoreIP(OMPBuilder.createTargetInit(Bld, IsSPMD, requiresFullRuntime()));
1052   IsInTargetMasterThreadRegion = IsSPMD;
1053   if (!IsSPMD)
1054     emitGenericVarsProlog(CGF, EST.Loc);
1055 }
1056 
1057 void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF,
1058                                           EntryFunctionState &EST,
1059                                           bool IsSPMD) {
1060   if (!IsSPMD)
1061     emitGenericVarsEpilog(CGF);
1062 
1063   CGBuilderTy &Bld = CGF.Builder;
1064   OMPBuilder.createTargetDeinit(Bld, IsSPMD, requiresFullRuntime());
1065 }
1066 
1067 void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D,
1068                                           StringRef ParentName,
1069                                           llvm::Function *&OutlinedFn,
1070                                           llvm::Constant *&OutlinedFnID,
1071                                           bool IsOffloadEntry,
1072                                           const RegionCodeGenTy &CodeGen) {
1073   ExecutionRuntimeModesRAII ModeRAII(
1074       CurrentExecutionMode, RequiresFullRuntime,
1075       CGM.getLangOpts().OpenMPCUDAForceFullRuntime ||
1076           !supportsLightweightRuntime(CGM.getContext(), D));
1077   EntryFunctionState EST;
1078 
1079   // Emit target region as a standalone region.
1080   class NVPTXPrePostActionTy : public PrePostActionTy {
1081     CGOpenMPRuntimeGPU &RT;
1082     CGOpenMPRuntimeGPU::EntryFunctionState &EST;
1083 
1084   public:
1085     NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT,
1086                          CGOpenMPRuntimeGPU::EntryFunctionState &EST)
1087         : RT(RT), EST(EST) {}
1088     void Enter(CodeGenFunction &CGF) override {
1089       RT.emitKernelInit(CGF, EST, /* IsSPMD */ true);
1090       // Skip target region initialization.
1091       RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
1092     }
1093     void Exit(CodeGenFunction &CGF) override {
1094       RT.clearLocThreadIdInsertPt(CGF);
1095       RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true);
1096     }
1097   } Action(*this, EST);
1098   CodeGen.setAction(Action);
1099   IsInTTDRegion = true;
1100   emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
1101                                    IsOffloadEntry, CodeGen);
1102   IsInTTDRegion = false;
1103 }
1104 
1105 // Create a unique global variable to indicate the execution mode of this target
1106 // region. The execution mode is either 'generic', or 'spmd' depending on the
1107 // target directive. This variable is picked up by the offload library to setup
1108 // the device appropriately before kernel launch. If the execution mode is
1109 // 'generic', the runtime reserves one warp for the master, otherwise, all
1110 // warps participate in parallel work.
1111 static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name,
1112                                      bool Mode) {
1113   auto *GVMode = new llvm::GlobalVariable(
1114       CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true,
1115       llvm::GlobalValue::WeakAnyLinkage,
1116       llvm::ConstantInt::get(CGM.Int8Ty, Mode ? OMP_TGT_EXEC_MODE_SPMD
1117                                               : OMP_TGT_EXEC_MODE_GENERIC),
1118       Twine(Name, "_exec_mode"));
1119   CGM.addCompilerUsedGlobal(GVMode);
1120 }
1121 
1122 void CGOpenMPRuntimeGPU::createOffloadEntry(llvm::Constant *ID,
1123                                               llvm::Constant *Addr,
1124                                               uint64_t Size, int32_t,
1125                                               llvm::GlobalValue::LinkageTypes) {
1126   // TODO: Add support for global variables on the device after declare target
1127   // support.
1128   if (!isa<llvm::Function>(Addr))
1129     return;
1130   llvm::Module &M = CGM.getModule();
1131   llvm::LLVMContext &Ctx = CGM.getLLVMContext();
1132 
1133   // Get "nvvm.annotations" metadata node
1134   llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations");
1135 
1136   llvm::Metadata *MDVals[] = {
1137       llvm::ConstantAsMetadata::get(Addr), llvm::MDString::get(Ctx, "kernel"),
1138       llvm::ConstantAsMetadata::get(
1139           llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))};
1140   // Append metadata to nvvm.annotations
1141   MD->addOperand(llvm::MDNode::get(Ctx, MDVals));
1142 }
1143 
1144 void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction(
1145     const OMPExecutableDirective &D, StringRef ParentName,
1146     llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
1147     bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
1148   if (!IsOffloadEntry) // Nothing to do.
1149     return;
1150 
1151   assert(!ParentName.empty() && "Invalid target region parent name!");
1152 
1153   bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D);
1154   if (Mode)
1155     emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
1156                    CodeGen);
1157   else
1158     emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
1159                       CodeGen);
1160 
1161   setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode);
1162 }
1163 
1164 namespace {
1165 LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE();
1166 /// Enum for accesseing the reserved_2 field of the ident_t struct.
1167 enum ModeFlagsTy : unsigned {
1168   /// Bit set to 1 when in SPMD mode.
1169   KMP_IDENT_SPMD_MODE = 0x01,
1170   /// Bit set to 1 when a simplified runtime is used.
1171   KMP_IDENT_SIMPLE_RT_MODE = 0x02,
1172   LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/KMP_IDENT_SIMPLE_RT_MODE)
1173 };
1174 
1175 /// Special mode Undefined. Is the combination of Non-SPMD mode + SimpleRuntime.
1176 static const ModeFlagsTy UndefinedMode =
1177     (~KMP_IDENT_SPMD_MODE) & KMP_IDENT_SIMPLE_RT_MODE;
1178 } // anonymous namespace
1179 
1180 unsigned CGOpenMPRuntimeGPU::getDefaultLocationReserved2Flags() const {
1181   switch (getExecutionMode()) {
1182   case EM_SPMD:
1183     if (requiresFullRuntime())
1184       return KMP_IDENT_SPMD_MODE & (~KMP_IDENT_SIMPLE_RT_MODE);
1185     return KMP_IDENT_SPMD_MODE | KMP_IDENT_SIMPLE_RT_MODE;
1186   case EM_NonSPMD:
1187     assert(requiresFullRuntime() && "Expected full runtime.");
1188     return (~KMP_IDENT_SPMD_MODE) & (~KMP_IDENT_SIMPLE_RT_MODE);
1189   case EM_Unknown:
1190     return UndefinedMode;
1191   }
1192   llvm_unreachable("Unknown flags are requested.");
1193 }
1194 
1195 CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM)
1196     : CGOpenMPRuntime(CGM, "_", "$") {
1197   if (!CGM.getLangOpts().OpenMPIsDevice)
1198     llvm_unreachable("OpenMP can only handle device code.");
1199 
1200   llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder();
1201   if (CGM.getLangOpts().OpenMPTargetNewRuntime) {
1202     OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTargetDebug,
1203                                 "__omp_rtl_debug_kind");
1204     OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTeamSubscription,
1205                                 "__omp_rtl_assume_teams_oversubscription");
1206     OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPThreadSubscription,
1207                                 "__omp_rtl_assume_threads_oversubscription");
1208   }
1209 }
1210 
1211 void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF,
1212                                               ProcBindKind ProcBind,
1213                                               SourceLocation Loc) {
1214   // Do nothing in case of SPMD mode and L0 parallel.
1215   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
1216     return;
1217 
1218   CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc);
1219 }
1220 
1221 void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF,
1222                                                 llvm::Value *NumThreads,
1223                                                 SourceLocation Loc) {
1224   // Nothing to do.
1225 }
1226 
1227 void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF,
1228                                               const Expr *NumTeams,
1229                                               const Expr *ThreadLimit,
1230                                               SourceLocation Loc) {}
1231 
1232 llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction(
1233     const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
1234     OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
1235   // Emit target region as a standalone region.
1236   class NVPTXPrePostActionTy : public PrePostActionTy {
1237     bool &IsInParallelRegion;
1238     bool PrevIsInParallelRegion;
1239 
1240   public:
1241     NVPTXPrePostActionTy(bool &IsInParallelRegion)
1242         : IsInParallelRegion(IsInParallelRegion) {}
1243     void Enter(CodeGenFunction &CGF) override {
1244       PrevIsInParallelRegion = IsInParallelRegion;
1245       IsInParallelRegion = true;
1246     }
1247     void Exit(CodeGenFunction &CGF) override {
1248       IsInParallelRegion = PrevIsInParallelRegion;
1249     }
1250   } Action(IsInParallelRegion);
1251   CodeGen.setAction(Action);
1252   bool PrevIsInTTDRegion = IsInTTDRegion;
1253   IsInTTDRegion = false;
1254   bool PrevIsInTargetMasterThreadRegion = IsInTargetMasterThreadRegion;
1255   IsInTargetMasterThreadRegion = false;
1256   auto *OutlinedFun =
1257       cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction(
1258           D, ThreadIDVar, InnermostKind, CodeGen));
1259   IsInTargetMasterThreadRegion = PrevIsInTargetMasterThreadRegion;
1260   IsInTTDRegion = PrevIsInTTDRegion;
1261   if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD &&
1262       !IsInParallelRegion) {
1263     llvm::Function *WrapperFun =
1264         createParallelDataSharingWrapper(OutlinedFun, D);
1265     WrapperFunctionsMap[OutlinedFun] = WrapperFun;
1266   }
1267 
1268   return OutlinedFun;
1269 }
1270 
1271 /// Get list of lastprivate variables from the teams distribute ... or
1272 /// teams {distribute ...} directives.
1273 static void
1274 getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D,
1275                              llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
1276   assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&
1277          "expected teams directive.");
1278   const OMPExecutableDirective *Dir = &D;
1279   if (!isOpenMPDistributeDirective(D.getDirectiveKind())) {
1280     if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild(
1281             Ctx,
1282             D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(
1283                 /*IgnoreCaptured=*/true))) {
1284       Dir = dyn_cast_or_null<OMPExecutableDirective>(S);
1285       if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind()))
1286         Dir = nullptr;
1287     }
1288   }
1289   if (!Dir)
1290     return;
1291   for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) {
1292     for (const Expr *E : C->getVarRefs())
1293       Vars.push_back(getPrivateItem(E));
1294   }
1295 }
1296 
1297 /// Get list of reduction variables from the teams ... directives.
1298 static void
1299 getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D,
1300                       llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
1301   assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&
1302          "expected teams directive.");
1303   for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
1304     for (const Expr *E : C->privates())
1305       Vars.push_back(getPrivateItem(E));
1306   }
1307 }
1308 
1309 llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction(
1310     const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
1311     OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
1312   SourceLocation Loc = D.getBeginLoc();
1313 
1314   const RecordDecl *GlobalizedRD = nullptr;
1315   llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions;
1316   llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
1317   unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size;
1318   // Globalize team reductions variable unconditionally in all modes.
1319   if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
1320     getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions);
1321   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
1322     getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions);
1323     if (!LastPrivatesReductions.empty()) {
1324       GlobalizedRD = ::buildRecordForGlobalizedVars(
1325           CGM.getContext(), llvm::None, LastPrivatesReductions,
1326           MappedDeclsFields, WarpSize);
1327     }
1328   } else if (!LastPrivatesReductions.empty()) {
1329     assert(!TeamAndReductions.first &&
1330            "Previous team declaration is not expected.");
1331     TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl();
1332     std::swap(TeamAndReductions.second, LastPrivatesReductions);
1333   }
1334 
1335   // Emit target region as a standalone region.
1336   class NVPTXPrePostActionTy : public PrePostActionTy {
1337     SourceLocation &Loc;
1338     const RecordDecl *GlobalizedRD;
1339     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
1340         &MappedDeclsFields;
1341 
1342   public:
1343     NVPTXPrePostActionTy(
1344         SourceLocation &Loc, const RecordDecl *GlobalizedRD,
1345         llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
1346             &MappedDeclsFields)
1347         : Loc(Loc), GlobalizedRD(GlobalizedRD),
1348           MappedDeclsFields(MappedDeclsFields) {}
1349     void Enter(CodeGenFunction &CGF) override {
1350       auto &Rt =
1351           static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1352       if (GlobalizedRD) {
1353         auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
1354         I->getSecond().MappedParams =
1355             std::make_unique<CodeGenFunction::OMPMapVars>();
1356         DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
1357         for (const auto &Pair : MappedDeclsFields) {
1358           assert(Pair.getFirst()->isCanonicalDecl() &&
1359                  "Expected canonical declaration");
1360           Data.insert(std::make_pair(Pair.getFirst(), MappedVarData()));
1361         }
1362       }
1363       Rt.emitGenericVarsProlog(CGF, Loc);
1364     }
1365     void Exit(CodeGenFunction &CGF) override {
1366       static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
1367           .emitGenericVarsEpilog(CGF);
1368     }
1369   } Action(Loc, GlobalizedRD, MappedDeclsFields);
1370   CodeGen.setAction(Action);
1371   llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction(
1372       D, ThreadIDVar, InnermostKind, CodeGen);
1373 
1374   return OutlinedFun;
1375 }
1376 
1377 void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF,
1378                                                  SourceLocation Loc,
1379                                                  bool WithSPMDCheck) {
1380   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic &&
1381       getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
1382     return;
1383 
1384   CGBuilderTy &Bld = CGF.Builder;
1385 
1386   const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
1387   if (I == FunctionGlobalizedDecls.end())
1388     return;
1389 
1390   for (auto &Rec : I->getSecond().LocalVarData) {
1391     const auto *VD = cast<VarDecl>(Rec.first);
1392     bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first);
1393     QualType VarTy = VD->getType();
1394 
1395     // Get the local allocation of a firstprivate variable before sharing
1396     llvm::Value *ParValue;
1397     if (EscapedParam) {
1398       LValue ParLVal =
1399           CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
1400       ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc);
1401     }
1402 
1403     // Allocate space for the variable to be globalized
1404     llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())};
1405     llvm::Instruction *VoidPtr =
1406         CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1407                                 CGM.getModule(), OMPRTL___kmpc_alloc_shared),
1408                             AllocArgs, VD->getName());
1409 
1410     // Cast the void pointer and get the address of the globalized variable.
1411     llvm::PointerType *VarPtrTy = CGF.ConvertTypeForMem(VarTy)->getPointerTo();
1412     llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
1413         VoidPtr, VarPtrTy, VD->getName() + "_on_stack");
1414     LValue VarAddr = CGF.MakeNaturalAlignAddrLValue(CastedVoidPtr, VarTy);
1415     Rec.second.PrivateAddr = VarAddr.getAddress(CGF);
1416     Rec.second.GlobalizedVal = VoidPtr;
1417 
1418     // Assign the local allocation to the newly globalized location.
1419     if (EscapedParam) {
1420       CGF.EmitStoreOfScalar(ParValue, VarAddr);
1421       I->getSecond().MappedParams->setVarAddr(CGF, VD, VarAddr.getAddress(CGF));
1422     }
1423     if (auto *DI = CGF.getDebugInfo())
1424       VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(VD->getLocation()));
1425   }
1426   for (const auto *VD : I->getSecond().EscapedVariableLengthDecls) {
1427     // Use actual memory size of the VLA object including the padding
1428     // for alignment purposes.
1429     llvm::Value *Size = CGF.getTypeSize(VD->getType());
1430     CharUnits Align = CGM.getContext().getDeclAlign(VD);
1431     Size = Bld.CreateNUWAdd(
1432         Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1));
1433     llvm::Value *AlignVal =
1434         llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity());
1435 
1436     Size = Bld.CreateUDiv(Size, AlignVal);
1437     Size = Bld.CreateNUWMul(Size, AlignVal);
1438 
1439     // Allocate space for this VLA object to be globalized.
1440     llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())};
1441     llvm::Instruction *VoidPtr =
1442         CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1443                                 CGM.getModule(), OMPRTL___kmpc_alloc_shared),
1444                             AllocArgs, VD->getName());
1445 
1446     I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(
1447         std::pair<llvm::Value *, llvm::Value *>(
1448             {VoidPtr, CGF.getTypeSize(VD->getType())}));
1449     LValue Base = CGF.MakeAddrLValue(VoidPtr, VD->getType(),
1450                                      CGM.getContext().getDeclAlign(VD),
1451                                      AlignmentSource::Decl);
1452     I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD),
1453                                             Base.getAddress(CGF));
1454   }
1455   I->getSecond().MappedParams->apply(CGF);
1456 }
1457 
1458 void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF,
1459                                                  bool WithSPMDCheck) {
1460   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic &&
1461       getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
1462     return;
1463 
1464   const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
1465   if (I != FunctionGlobalizedDecls.end()) {
1466     // Deallocate the memory for each globalized VLA object
1467     for (auto AddrSizePair :
1468          llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) {
1469       CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1470                               CGM.getModule(), OMPRTL___kmpc_free_shared),
1471                           {AddrSizePair.first, AddrSizePair.second});
1472     }
1473     // Deallocate the memory for each globalized value
1474     for (auto &Rec : llvm::reverse(I->getSecond().LocalVarData)) {
1475       const auto *VD = cast<VarDecl>(Rec.first);
1476       I->getSecond().MappedParams->restore(CGF);
1477 
1478       llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal,
1479                                  CGF.getTypeSize(VD->getType())};
1480       CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1481                               CGM.getModule(), OMPRTL___kmpc_free_shared),
1482                           FreeArgs);
1483     }
1484   }
1485 }
1486 
1487 void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF,
1488                                          const OMPExecutableDirective &D,
1489                                          SourceLocation Loc,
1490                                          llvm::Function *OutlinedFn,
1491                                          ArrayRef<llvm::Value *> CapturedVars) {
1492   if (!CGF.HaveInsertPoint())
1493     return;
1494 
1495   Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
1496                                                       /*Name=*/".zero.addr");
1497   CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr);
1498   llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
1499   OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer());
1500   OutlinedFnArgs.push_back(ZeroAddr.getPointer());
1501   OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
1502   emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs);
1503 }
1504 
1505 void CGOpenMPRuntimeGPU::emitParallelCall(CodeGenFunction &CGF,
1506                                           SourceLocation Loc,
1507                                           llvm::Function *OutlinedFn,
1508                                           ArrayRef<llvm::Value *> CapturedVars,
1509                                           const Expr *IfCond,
1510                                           llvm::Value *NumThreads) {
1511   if (!CGF.HaveInsertPoint())
1512     return;
1513 
1514   auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond,
1515                         NumThreads](CodeGenFunction &CGF,
1516                                     PrePostActionTy &Action) {
1517     CGBuilderTy &Bld = CGF.Builder;
1518     llvm::Value *NumThreadsVal = NumThreads;
1519     llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn];
1520     llvm::Value *ID = llvm::ConstantPointerNull::get(CGM.Int8PtrTy);
1521     if (WFn)
1522       ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy);
1523     llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(OutlinedFn, CGM.Int8PtrTy);
1524 
1525     // Create a private scope that will globalize the arguments
1526     // passed from the outside of the target region.
1527     // TODO: Is that needed?
1528     CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF);
1529 
1530     Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca(
1531         llvm::ArrayType::get(CGM.VoidPtrTy, CapturedVars.size()),
1532         "captured_vars_addrs");
1533     // There's something to share.
1534     if (!CapturedVars.empty()) {
1535       // Prepare for parallel region. Indicate the outlined function.
1536       ASTContext &Ctx = CGF.getContext();
1537       unsigned Idx = 0;
1538       for (llvm::Value *V : CapturedVars) {
1539         Address Dst = Bld.CreateConstArrayGEP(CapturedVarsAddrs, Idx);
1540         llvm::Value *PtrV;
1541         if (V->getType()->isIntegerTy())
1542           PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy);
1543         else
1544           PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy);
1545         CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false,
1546                               Ctx.getPointerType(Ctx.VoidPtrTy));
1547         ++Idx;
1548       }
1549     }
1550 
1551     llvm::Value *IfCondVal = nullptr;
1552     if (IfCond)
1553       IfCondVal = Bld.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.Int32Ty,
1554                                     /* isSigned */ false);
1555     else
1556       IfCondVal = llvm::ConstantInt::get(CGF.Int32Ty, 1);
1557 
1558     if (!NumThreadsVal)
1559       NumThreadsVal = llvm::ConstantInt::get(CGF.Int32Ty, -1);
1560     else
1561       NumThreadsVal = Bld.CreateZExtOrTrunc(NumThreadsVal, CGF.Int32Ty),
1562 
1563       assert(IfCondVal && "Expected a value");
1564     llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
1565     llvm::Value *Args[] = {
1566         RTLoc,
1567         getThreadID(CGF, Loc),
1568         IfCondVal,
1569         NumThreadsVal,
1570         llvm::ConstantInt::get(CGF.Int32Ty, -1),
1571         FnPtr,
1572         ID,
1573         Bld.CreateBitOrPointerCast(CapturedVarsAddrs.getPointer(),
1574                                    CGF.VoidPtrPtrTy),
1575         llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())};
1576     CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1577                             CGM.getModule(), OMPRTL___kmpc_parallel_51),
1578                         Args);
1579   };
1580 
1581   RegionCodeGenTy RCG(ParallelGen);
1582   RCG(CGF);
1583 }
1584 
1585 void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) {
1586   // Always emit simple barriers!
1587   if (!CGF.HaveInsertPoint())
1588     return;
1589   // Build call __kmpc_barrier_simple_spmd(nullptr, 0);
1590   // This function does not use parameters, so we can emit just default values.
1591   llvm::Value *Args[] = {
1592       llvm::ConstantPointerNull::get(
1593           cast<llvm::PointerType>(getIdentTyPointerTy())),
1594       llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)};
1595   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1596                           CGM.getModule(), OMPRTL___kmpc_barrier_simple_spmd),
1597                       Args);
1598 }
1599 
1600 void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF,
1601                                            SourceLocation Loc,
1602                                            OpenMPDirectiveKind Kind, bool,
1603                                            bool) {
1604   // Always emit simple barriers!
1605   if (!CGF.HaveInsertPoint())
1606     return;
1607   // Build call __kmpc_cancel_barrier(loc, thread_id);
1608   unsigned Flags = getDefaultFlagsForBarriers(Kind);
1609   llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags),
1610                          getThreadID(CGF, Loc)};
1611 
1612   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1613                           CGM.getModule(), OMPRTL___kmpc_barrier),
1614                       Args);
1615 }
1616 
1617 void CGOpenMPRuntimeGPU::emitCriticalRegion(
1618     CodeGenFunction &CGF, StringRef CriticalName,
1619     const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc,
1620     const Expr *Hint) {
1621   llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop");
1622   llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test");
1623   llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync");
1624   llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body");
1625   llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit");
1626 
1627   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1628 
1629   // Get the mask of active threads in the warp.
1630   llvm::Value *Mask = CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1631       CGM.getModule(), OMPRTL___kmpc_warp_active_thread_mask));
1632   // Fetch team-local id of the thread.
1633   llvm::Value *ThreadID = RT.getGPUThreadID(CGF);
1634 
1635   // Get the width of the team.
1636   llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF);
1637 
1638   // Initialize the counter variable for the loop.
1639   QualType Int32Ty =
1640       CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0);
1641   Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter");
1642   LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty);
1643   CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal,
1644                         /*isInit=*/true);
1645 
1646   // Block checks if loop counter exceeds upper bound.
1647   CGF.EmitBlock(LoopBB);
1648   llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
1649   llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth);
1650   CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB);
1651 
1652   // Block tests which single thread should execute region, and which threads
1653   // should go straight to synchronisation point.
1654   CGF.EmitBlock(TestBB);
1655   CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
1656   llvm::Value *CmpThreadToCounter =
1657       CGF.Builder.CreateICmpEQ(ThreadID, CounterVal);
1658   CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB);
1659 
1660   // Block emits the body of the critical region.
1661   CGF.EmitBlock(BodyBB);
1662 
1663   // Output the critical statement.
1664   CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc,
1665                                       Hint);
1666 
1667   // After the body surrounded by the critical region, the single executing
1668   // thread will jump to the synchronisation point.
1669   // Block waits for all threads in current team to finish then increments the
1670   // counter variable and returns to the loop.
1671   CGF.EmitBlock(SyncBB);
1672   // Reconverge active threads in the warp.
1673   (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1674                                 CGM.getModule(), OMPRTL___kmpc_syncwarp),
1675                             Mask);
1676 
1677   llvm::Value *IncCounterVal =
1678       CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1));
1679   CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal);
1680   CGF.EmitBranch(LoopBB);
1681 
1682   // Block that is reached when  all threads in the team complete the region.
1683   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1684 }
1685 
1686 /// Cast value to the specified type.
1687 static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val,
1688                                     QualType ValTy, QualType CastTy,
1689                                     SourceLocation Loc) {
1690   assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() &&
1691          "Cast type must sized.");
1692   assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() &&
1693          "Val type must sized.");
1694   llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy);
1695   if (ValTy == CastTy)
1696     return Val;
1697   if (CGF.getContext().getTypeSizeInChars(ValTy) ==
1698       CGF.getContext().getTypeSizeInChars(CastTy))
1699     return CGF.Builder.CreateBitCast(Val, LLVMCastTy);
1700   if (CastTy->isIntegerType() && ValTy->isIntegerType())
1701     return CGF.Builder.CreateIntCast(Val, LLVMCastTy,
1702                                      CastTy->hasSignedIntegerRepresentation());
1703   Address CastItem = CGF.CreateMemTemp(CastTy);
1704   Address ValCastItem = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
1705       CastItem, Val->getType()->getPointerTo(CastItem.getAddressSpace()));
1706   CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy,
1707                         LValueBaseInfo(AlignmentSource::Type),
1708                         TBAAAccessInfo());
1709   return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc,
1710                               LValueBaseInfo(AlignmentSource::Type),
1711                               TBAAAccessInfo());
1712 }
1713 
1714 /// This function creates calls to one of two shuffle functions to copy
1715 /// variables between lanes in a warp.
1716 static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF,
1717                                                  llvm::Value *Elem,
1718                                                  QualType ElemType,
1719                                                  llvm::Value *Offset,
1720                                                  SourceLocation Loc) {
1721   CodeGenModule &CGM = CGF.CGM;
1722   CGBuilderTy &Bld = CGF.Builder;
1723   CGOpenMPRuntimeGPU &RT =
1724       *(static_cast<CGOpenMPRuntimeGPU *>(&CGM.getOpenMPRuntime()));
1725   llvm::OpenMPIRBuilder &OMPBuilder = RT.getOMPBuilder();
1726 
1727   CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType);
1728   assert(Size.getQuantity() <= 8 &&
1729          "Unsupported bitwidth in shuffle instruction.");
1730 
1731   RuntimeFunction ShuffleFn = Size.getQuantity() <= 4
1732                                   ? OMPRTL___kmpc_shuffle_int32
1733                                   : OMPRTL___kmpc_shuffle_int64;
1734 
1735   // Cast all types to 32- or 64-bit values before calling shuffle routines.
1736   QualType CastTy = CGF.getContext().getIntTypeForBitwidth(
1737       Size.getQuantity() <= 4 ? 32 : 64, /*Signed=*/1);
1738   llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc);
1739   llvm::Value *WarpSize =
1740       Bld.CreateIntCast(RT.getGPUWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true);
1741 
1742   llvm::Value *ShuffledVal = CGF.EmitRuntimeCall(
1743       OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), ShuffleFn),
1744       {ElemCast, Offset, WarpSize});
1745 
1746   return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc);
1747 }
1748 
1749 static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr,
1750                             Address DestAddr, QualType ElemType,
1751                             llvm::Value *Offset, SourceLocation Loc) {
1752   CGBuilderTy &Bld = CGF.Builder;
1753 
1754   CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType);
1755   // Create the loop over the big sized data.
1756   // ptr = (void*)Elem;
1757   // ptrEnd = (void*) Elem + 1;
1758   // Step = 8;
1759   // while (ptr + Step < ptrEnd)
1760   //   shuffle((int64_t)*ptr);
1761   // Step = 4;
1762   // while (ptr + Step < ptrEnd)
1763   //   shuffle((int32_t)*ptr);
1764   // ...
1765   Address ElemPtr = DestAddr;
1766   Address Ptr = SrcAddr;
1767   Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast(
1768       Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy);
1769   for (int IntSize = 8; IntSize >= 1; IntSize /= 2) {
1770     if (Size < CharUnits::fromQuantity(IntSize))
1771       continue;
1772     QualType IntType = CGF.getContext().getIntTypeForBitwidth(
1773         CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)),
1774         /*Signed=*/1);
1775     llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType);
1776     Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo());
1777     ElemPtr =
1778         Bld.CreatePointerBitCastOrAddrSpaceCast(ElemPtr, IntTy->getPointerTo());
1779     if (Size.getQuantity() / IntSize > 1) {
1780       llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond");
1781       llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then");
1782       llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit");
1783       llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock();
1784       CGF.EmitBlock(PreCondBB);
1785       llvm::PHINode *PhiSrc =
1786           Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2);
1787       PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB);
1788       llvm::PHINode *PhiDest =
1789           Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2);
1790       PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB);
1791       Ptr = Address(PhiSrc, Ptr.getAlignment());
1792       ElemPtr = Address(PhiDest, ElemPtr.getAlignment());
1793       llvm::Value *PtrDiff = Bld.CreatePtrDiff(
1794           PtrEnd.getPointer(), Bld.CreatePointerBitCastOrAddrSpaceCast(
1795                                    Ptr.getPointer(), CGF.VoidPtrTy));
1796       Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)),
1797                        ThenBB, ExitBB);
1798       CGF.EmitBlock(ThenBB);
1799       llvm::Value *Res = createRuntimeShuffleFunction(
1800           CGF,
1801           CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc,
1802                                LValueBaseInfo(AlignmentSource::Type),
1803                                TBAAAccessInfo()),
1804           IntType, Offset, Loc);
1805       CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType,
1806                             LValueBaseInfo(AlignmentSource::Type),
1807                             TBAAAccessInfo());
1808       Address LocalPtr = Bld.CreateConstGEP(Ptr, 1);
1809       Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1);
1810       PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB);
1811       PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB);
1812       CGF.EmitBranch(PreCondBB);
1813       CGF.EmitBlock(ExitBB);
1814     } else {
1815       llvm::Value *Res = createRuntimeShuffleFunction(
1816           CGF,
1817           CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc,
1818                                LValueBaseInfo(AlignmentSource::Type),
1819                                TBAAAccessInfo()),
1820           IntType, Offset, Loc);
1821       CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType,
1822                             LValueBaseInfo(AlignmentSource::Type),
1823                             TBAAAccessInfo());
1824       Ptr = Bld.CreateConstGEP(Ptr, 1);
1825       ElemPtr = Bld.CreateConstGEP(ElemPtr, 1);
1826     }
1827     Size = Size % IntSize;
1828   }
1829 }
1830 
1831 namespace {
1832 enum CopyAction : unsigned {
1833   // RemoteLaneToThread: Copy over a Reduce list from a remote lane in
1834   // the warp using shuffle instructions.
1835   RemoteLaneToThread,
1836   // ThreadCopy: Make a copy of a Reduce list on the thread's stack.
1837   ThreadCopy,
1838   // ThreadToScratchpad: Copy a team-reduced array to the scratchpad.
1839   ThreadToScratchpad,
1840   // ScratchpadToThread: Copy from a scratchpad array in global memory
1841   // containing team-reduced data to a thread's stack.
1842   ScratchpadToThread,
1843 };
1844 } // namespace
1845 
1846 struct CopyOptionsTy {
1847   llvm::Value *RemoteLaneOffset;
1848   llvm::Value *ScratchpadIndex;
1849   llvm::Value *ScratchpadWidth;
1850 };
1851 
1852 /// Emit instructions to copy a Reduce list, which contains partially
1853 /// aggregated values, in the specified direction.
1854 static void emitReductionListCopy(
1855     CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy,
1856     ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase,
1857     CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) {
1858 
1859   CodeGenModule &CGM = CGF.CGM;
1860   ASTContext &C = CGM.getContext();
1861   CGBuilderTy &Bld = CGF.Builder;
1862 
1863   llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset;
1864   llvm::Value *ScratchpadIndex = CopyOptions.ScratchpadIndex;
1865   llvm::Value *ScratchpadWidth = CopyOptions.ScratchpadWidth;
1866 
1867   // Iterates, element-by-element, through the source Reduce list and
1868   // make a copy.
1869   unsigned Idx = 0;
1870   unsigned Size = Privates.size();
1871   for (const Expr *Private : Privates) {
1872     Address SrcElementAddr = Address::invalid();
1873     Address DestElementAddr = Address::invalid();
1874     Address DestElementPtrAddr = Address::invalid();
1875     // Should we shuffle in an element from a remote lane?
1876     bool ShuffleInElement = false;
1877     // Set to true to update the pointer in the dest Reduce list to a
1878     // newly created element.
1879     bool UpdateDestListPtr = false;
1880     // Increment the src or dest pointer to the scratchpad, for each
1881     // new element.
1882     bool IncrScratchpadSrc = false;
1883     bool IncrScratchpadDest = false;
1884 
1885     switch (Action) {
1886     case RemoteLaneToThread: {
1887       // Step 1.1: Get the address for the src element in the Reduce list.
1888       Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
1889       SrcElementAddr = CGF.EmitLoadOfPointer(
1890           SrcElementPtrAddr,
1891           C.getPointerType(Private->getType())->castAs<PointerType>());
1892 
1893       // Step 1.2: Create a temporary to store the element in the destination
1894       // Reduce list.
1895       DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
1896       DestElementAddr =
1897           CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element");
1898       ShuffleInElement = true;
1899       UpdateDestListPtr = true;
1900       break;
1901     }
1902     case ThreadCopy: {
1903       // Step 1.1: Get the address for the src element in the Reduce list.
1904       Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
1905       SrcElementAddr = CGF.EmitLoadOfPointer(
1906           SrcElementPtrAddr,
1907           C.getPointerType(Private->getType())->castAs<PointerType>());
1908 
1909       // Step 1.2: Get the address for dest element.  The destination
1910       // element has already been created on the thread's stack.
1911       DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
1912       DestElementAddr = CGF.EmitLoadOfPointer(
1913           DestElementPtrAddr,
1914           C.getPointerType(Private->getType())->castAs<PointerType>());
1915       break;
1916     }
1917     case ThreadToScratchpad: {
1918       // Step 1.1: Get the address for the src element in the Reduce list.
1919       Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
1920       SrcElementAddr = CGF.EmitLoadOfPointer(
1921           SrcElementPtrAddr,
1922           C.getPointerType(Private->getType())->castAs<PointerType>());
1923 
1924       // Step 1.2: Get the address for dest element:
1925       // address = base + index * ElementSizeInChars.
1926       llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
1927       llvm::Value *CurrentOffset =
1928           Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex);
1929       llvm::Value *ScratchPadElemAbsolutePtrVal =
1930           Bld.CreateNUWAdd(DestBase.getPointer(), CurrentOffset);
1931       ScratchPadElemAbsolutePtrVal =
1932           Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy);
1933       DestElementAddr = Address(ScratchPadElemAbsolutePtrVal,
1934                                 C.getTypeAlignInChars(Private->getType()));
1935       IncrScratchpadDest = true;
1936       break;
1937     }
1938     case ScratchpadToThread: {
1939       // Step 1.1: Get the address for the src element in the scratchpad.
1940       // address = base + index * ElementSizeInChars.
1941       llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
1942       llvm::Value *CurrentOffset =
1943           Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex);
1944       llvm::Value *ScratchPadElemAbsolutePtrVal =
1945           Bld.CreateNUWAdd(SrcBase.getPointer(), CurrentOffset);
1946       ScratchPadElemAbsolutePtrVal =
1947           Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy);
1948       SrcElementAddr = Address(ScratchPadElemAbsolutePtrVal,
1949                                C.getTypeAlignInChars(Private->getType()));
1950       IncrScratchpadSrc = true;
1951 
1952       // Step 1.2: Create a temporary to store the element in the destination
1953       // Reduce list.
1954       DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
1955       DestElementAddr =
1956           CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element");
1957       UpdateDestListPtr = true;
1958       break;
1959     }
1960     }
1961 
1962     // Regardless of src and dest of copy, we emit the load of src
1963     // element as this is required in all directions
1964     SrcElementAddr = Bld.CreateElementBitCast(
1965         SrcElementAddr, CGF.ConvertTypeForMem(Private->getType()));
1966     DestElementAddr = Bld.CreateElementBitCast(DestElementAddr,
1967                                                SrcElementAddr.getElementType());
1968 
1969     // Now that all active lanes have read the element in the
1970     // Reduce list, shuffle over the value from the remote lane.
1971     if (ShuffleInElement) {
1972       shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(),
1973                       RemoteLaneOffset, Private->getExprLoc());
1974     } else {
1975       switch (CGF.getEvaluationKind(Private->getType())) {
1976       case TEK_Scalar: {
1977         llvm::Value *Elem = CGF.EmitLoadOfScalar(
1978             SrcElementAddr, /*Volatile=*/false, Private->getType(),
1979             Private->getExprLoc(), LValueBaseInfo(AlignmentSource::Type),
1980             TBAAAccessInfo());
1981         // Store the source element value to the dest element address.
1982         CGF.EmitStoreOfScalar(
1983             Elem, DestElementAddr, /*Volatile=*/false, Private->getType(),
1984             LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
1985         break;
1986       }
1987       case TEK_Complex: {
1988         CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex(
1989             CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
1990             Private->getExprLoc());
1991         CGF.EmitStoreOfComplex(
1992             Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
1993             /*isInit=*/false);
1994         break;
1995       }
1996       case TEK_Aggregate:
1997         CGF.EmitAggregateCopy(
1998             CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
1999             CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
2000             Private->getType(), AggValueSlot::DoesNotOverlap);
2001         break;
2002       }
2003     }
2004 
2005     // Step 3.1: Modify reference in dest Reduce list as needed.
2006     // Modifying the reference in Reduce list to point to the newly
2007     // created element.  The element is live in the current function
2008     // scope and that of functions it invokes (i.e., reduce_function).
2009     // RemoteReduceData[i] = (void*)&RemoteElem
2010     if (UpdateDestListPtr) {
2011       CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast(
2012                                 DestElementAddr.getPointer(), CGF.VoidPtrTy),
2013                             DestElementPtrAddr, /*Volatile=*/false,
2014                             C.VoidPtrTy);
2015     }
2016 
2017     // Step 4.1: Increment SrcBase/DestBase so that it points to the starting
2018     // address of the next element in scratchpad memory, unless we're currently
2019     // processing the last one.  Memory alignment is also taken care of here.
2020     if ((IncrScratchpadDest || IncrScratchpadSrc) && (Idx + 1 < Size)) {
2021       llvm::Value *ScratchpadBasePtr =
2022           IncrScratchpadDest ? DestBase.getPointer() : SrcBase.getPointer();
2023       llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
2024       ScratchpadBasePtr = Bld.CreateNUWAdd(
2025           ScratchpadBasePtr,
2026           Bld.CreateNUWMul(ScratchpadWidth, ElementSizeInChars));
2027 
2028       // Take care of global memory alignment for performance
2029       ScratchpadBasePtr = Bld.CreateNUWSub(
2030           ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1));
2031       ScratchpadBasePtr = Bld.CreateUDiv(
2032           ScratchpadBasePtr,
2033           llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment));
2034       ScratchpadBasePtr = Bld.CreateNUWAdd(
2035           ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1));
2036       ScratchpadBasePtr = Bld.CreateNUWMul(
2037           ScratchpadBasePtr,
2038           llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment));
2039 
2040       if (IncrScratchpadDest)
2041         DestBase = Address(ScratchpadBasePtr, CGF.getPointerAlign());
2042       else /* IncrScratchpadSrc = true */
2043         SrcBase = Address(ScratchpadBasePtr, CGF.getPointerAlign());
2044     }
2045 
2046     ++Idx;
2047   }
2048 }
2049 
2050 /// This function emits a helper that gathers Reduce lists from the first
2051 /// lane of every active warp to lanes in the first warp.
2052 ///
2053 /// void inter_warp_copy_func(void* reduce_data, num_warps)
2054 ///   shared smem[warp_size];
2055 ///   For all data entries D in reduce_data:
2056 ///     sync
2057 ///     If (I am the first lane in each warp)
2058 ///       Copy my local D to smem[warp_id]
2059 ///     sync
2060 ///     if (I am the first warp)
2061 ///       Copy smem[thread_id] to my local D
2062 static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM,
2063                                               ArrayRef<const Expr *> Privates,
2064                                               QualType ReductionArrayTy,
2065                                               SourceLocation Loc) {
2066   ASTContext &C = CGM.getContext();
2067   llvm::Module &M = CGM.getModule();
2068 
2069   // ReduceList: thread local Reduce list.
2070   // At the stage of the computation when this function is called, partially
2071   // aggregated values reside in the first lane of every active warp.
2072   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2073                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2074   // NumWarps: number of warps active in the parallel region.  This could
2075   // be smaller than 32 (max warps in a CTA) for partial block reduction.
2076   ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2077                                 C.getIntTypeForBitwidth(32, /* Signed */ true),
2078                                 ImplicitParamDecl::Other);
2079   FunctionArgList Args;
2080   Args.push_back(&ReduceListArg);
2081   Args.push_back(&NumWarpsArg);
2082 
2083   const CGFunctionInfo &CGFI =
2084       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2085   auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI),
2086                                     llvm::GlobalValue::InternalLinkage,
2087                                     "_omp_reduction_inter_warp_copy_func", &M);
2088   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2089   Fn->setDoesNotRecurse();
2090   CodeGenFunction CGF(CGM);
2091   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2092 
2093   CGBuilderTy &Bld = CGF.Builder;
2094 
2095   // This array is used as a medium to transfer, one reduce element at a time,
2096   // the data from the first lane of every warp to lanes in the first warp
2097   // in order to perform the final step of a reduction in a parallel region
2098   // (reduction across warps).  The array is placed in NVPTX __shared__ memory
2099   // for reduced latency, as well as to have a distinct copy for concurrently
2100   // executing target regions.  The array is declared with common linkage so
2101   // as to be shared across compilation units.
2102   StringRef TransferMediumName =
2103       "__openmp_nvptx_data_transfer_temporary_storage";
2104   llvm::GlobalVariable *TransferMedium =
2105       M.getGlobalVariable(TransferMediumName);
2106   unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size;
2107   if (!TransferMedium) {
2108     auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize);
2109     unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared);
2110     TransferMedium = new llvm::GlobalVariable(
2111         M, Ty, /*isConstant=*/false, llvm::GlobalVariable::WeakAnyLinkage,
2112         llvm::UndefValue::get(Ty), TransferMediumName,
2113         /*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal,
2114         SharedAddressSpace);
2115     CGM.addCompilerUsedGlobal(TransferMedium);
2116   }
2117 
2118   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
2119   // Get the CUDA thread id of the current OpenMP thread on the GPU.
2120   llvm::Value *ThreadID = RT.getGPUThreadID(CGF);
2121   // nvptx_lane_id = nvptx_id % warpsize
2122   llvm::Value *LaneID = getNVPTXLaneID(CGF);
2123   // nvptx_warp_id = nvptx_id / warpsize
2124   llvm::Value *WarpID = getNVPTXWarpID(CGF);
2125 
2126   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2127   Address LocalReduceList(
2128       Bld.CreatePointerBitCastOrAddrSpaceCast(
2129           CGF.EmitLoadOfScalar(
2130               AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc,
2131               LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()),
2132           CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
2133       CGF.getPointerAlign());
2134 
2135   unsigned Idx = 0;
2136   for (const Expr *Private : Privates) {
2137     //
2138     // Warp master copies reduce element to transfer medium in __shared__
2139     // memory.
2140     //
2141     unsigned RealTySize =
2142         C.getTypeSizeInChars(Private->getType())
2143             .alignTo(C.getTypeAlignInChars(Private->getType()))
2144             .getQuantity();
2145     for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /=2) {
2146       unsigned NumIters = RealTySize / TySize;
2147       if (NumIters == 0)
2148         continue;
2149       QualType CType = C.getIntTypeForBitwidth(
2150           C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1);
2151       llvm::Type *CopyType = CGF.ConvertTypeForMem(CType);
2152       CharUnits Align = CharUnits::fromQuantity(TySize);
2153       llvm::Value *Cnt = nullptr;
2154       Address CntAddr = Address::invalid();
2155       llvm::BasicBlock *PrecondBB = nullptr;
2156       llvm::BasicBlock *ExitBB = nullptr;
2157       if (NumIters > 1) {
2158         CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr");
2159         CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr,
2160                               /*Volatile=*/false, C.IntTy);
2161         PrecondBB = CGF.createBasicBlock("precond");
2162         ExitBB = CGF.createBasicBlock("exit");
2163         llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body");
2164         // There is no need to emit line number for unconditional branch.
2165         (void)ApplyDebugLocation::CreateEmpty(CGF);
2166         CGF.EmitBlock(PrecondBB);
2167         Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc);
2168         llvm::Value *Cmp =
2169             Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters));
2170         Bld.CreateCondBr(Cmp, BodyBB, ExitBB);
2171         CGF.EmitBlock(BodyBB);
2172       }
2173       // kmpc_barrier.
2174       CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
2175                                              /*EmitChecks=*/false,
2176                                              /*ForceSimpleCall=*/true);
2177       llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then");
2178       llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else");
2179       llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont");
2180 
2181       // if (lane_id == 0)
2182       llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master");
2183       Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB);
2184       CGF.EmitBlock(ThenBB);
2185 
2186       // Reduce element = LocalReduceList[i]
2187       Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
2188       llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
2189           ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
2190       // elemptr = ((CopyType*)(elemptrptr)) + I
2191       Address ElemPtr = Address(ElemPtrPtr, Align);
2192       ElemPtr = Bld.CreateElementBitCast(ElemPtr, CopyType);
2193       if (NumIters > 1) {
2194         ElemPtr = Address(Bld.CreateGEP(ElemPtr.getElementType(),
2195                                         ElemPtr.getPointer(), Cnt),
2196                           ElemPtr.getAlignment());
2197       }
2198 
2199       // Get pointer to location in transfer medium.
2200       // MediumPtr = &medium[warp_id]
2201       llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP(
2202           TransferMedium->getValueType(), TransferMedium,
2203           {llvm::Constant::getNullValue(CGM.Int64Ty), WarpID});
2204       Address MediumPtr(MediumPtrVal, Align);
2205       // Casting to actual data type.
2206       // MediumPtr = (CopyType*)MediumPtrAddr;
2207       MediumPtr = Bld.CreateElementBitCast(MediumPtr, CopyType);
2208 
2209       // elem = *elemptr
2210       //*MediumPtr = elem
2211       llvm::Value *Elem = CGF.EmitLoadOfScalar(
2212           ElemPtr, /*Volatile=*/false, CType, Loc,
2213           LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
2214       // Store the source element value to the dest element address.
2215       CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType,
2216                             LValueBaseInfo(AlignmentSource::Type),
2217                             TBAAAccessInfo());
2218 
2219       Bld.CreateBr(MergeBB);
2220 
2221       CGF.EmitBlock(ElseBB);
2222       Bld.CreateBr(MergeBB);
2223 
2224       CGF.EmitBlock(MergeBB);
2225 
2226       // kmpc_barrier.
2227       CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
2228                                              /*EmitChecks=*/false,
2229                                              /*ForceSimpleCall=*/true);
2230 
2231       //
2232       // Warp 0 copies reduce element from transfer medium.
2233       //
2234       llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then");
2235       llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else");
2236       llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont");
2237 
2238       Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg);
2239       llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar(
2240           AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc);
2241 
2242       // Up to 32 threads in warp 0 are active.
2243       llvm::Value *IsActiveThread =
2244           Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread");
2245       Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB);
2246 
2247       CGF.EmitBlock(W0ThenBB);
2248 
2249       // SrcMediumPtr = &medium[tid]
2250       llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP(
2251           TransferMedium->getValueType(), TransferMedium,
2252           {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID});
2253       Address SrcMediumPtr(SrcMediumPtrVal, Align);
2254       // SrcMediumVal = *SrcMediumPtr;
2255       SrcMediumPtr = Bld.CreateElementBitCast(SrcMediumPtr, CopyType);
2256 
2257       // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I
2258       Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
2259       llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar(
2260           TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc);
2261       Address TargetElemPtr = Address(TargetElemPtrVal, Align);
2262       TargetElemPtr = Bld.CreateElementBitCast(TargetElemPtr, CopyType);
2263       if (NumIters > 1) {
2264         TargetElemPtr = Address(Bld.CreateGEP(TargetElemPtr.getElementType(),
2265                                               TargetElemPtr.getPointer(), Cnt),
2266                                 TargetElemPtr.getAlignment());
2267       }
2268 
2269       // *TargetElemPtr = SrcMediumVal;
2270       llvm::Value *SrcMediumValue =
2271           CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc);
2272       CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false,
2273                             CType);
2274       Bld.CreateBr(W0MergeBB);
2275 
2276       CGF.EmitBlock(W0ElseBB);
2277       Bld.CreateBr(W0MergeBB);
2278 
2279       CGF.EmitBlock(W0MergeBB);
2280 
2281       if (NumIters > 1) {
2282         Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1));
2283         CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy);
2284         CGF.EmitBranch(PrecondBB);
2285         (void)ApplyDebugLocation::CreateEmpty(CGF);
2286         CGF.EmitBlock(ExitBB);
2287       }
2288       RealTySize %= TySize;
2289     }
2290     ++Idx;
2291   }
2292 
2293   CGF.FinishFunction();
2294   return Fn;
2295 }
2296 
2297 /// Emit a helper that reduces data across two OpenMP threads (lanes)
2298 /// in the same warp.  It uses shuffle instructions to copy over data from
2299 /// a remote lane's stack.  The reduction algorithm performed is specified
2300 /// by the fourth parameter.
2301 ///
2302 /// Algorithm Versions.
2303 /// Full Warp Reduce (argument value 0):
2304 ///   This algorithm assumes that all 32 lanes are active and gathers
2305 ///   data from these 32 lanes, producing a single resultant value.
2306 /// Contiguous Partial Warp Reduce (argument value 1):
2307 ///   This algorithm assumes that only a *contiguous* subset of lanes
2308 ///   are active.  This happens for the last warp in a parallel region
2309 ///   when the user specified num_threads is not an integer multiple of
2310 ///   32.  This contiguous subset always starts with the zeroth lane.
2311 /// Partial Warp Reduce (argument value 2):
2312 ///   This algorithm gathers data from any number of lanes at any position.
2313 /// All reduced values are stored in the lowest possible lane.  The set
2314 /// of problems every algorithm addresses is a super set of those
2315 /// addressable by algorithms with a lower version number.  Overhead
2316 /// increases as algorithm version increases.
2317 ///
2318 /// Terminology
2319 /// Reduce element:
2320 ///   Reduce element refers to the individual data field with primitive
2321 ///   data types to be combined and reduced across threads.
2322 /// Reduce list:
2323 ///   Reduce list refers to a collection of local, thread-private
2324 ///   reduce elements.
2325 /// Remote Reduce list:
2326 ///   Remote Reduce list refers to a collection of remote (relative to
2327 ///   the current thread) reduce elements.
2328 ///
2329 /// We distinguish between three states of threads that are important to
2330 /// the implementation of this function.
2331 /// Alive threads:
2332 ///   Threads in a warp executing the SIMT instruction, as distinguished from
2333 ///   threads that are inactive due to divergent control flow.
2334 /// Active threads:
2335 ///   The minimal set of threads that has to be alive upon entry to this
2336 ///   function.  The computation is correct iff active threads are alive.
2337 ///   Some threads are alive but they are not active because they do not
2338 ///   contribute to the computation in any useful manner.  Turning them off
2339 ///   may introduce control flow overheads without any tangible benefits.
2340 /// Effective threads:
2341 ///   In order to comply with the argument requirements of the shuffle
2342 ///   function, we must keep all lanes holding data alive.  But at most
2343 ///   half of them perform value aggregation; we refer to this half of
2344 ///   threads as effective. The other half is simply handing off their
2345 ///   data.
2346 ///
2347 /// Procedure
2348 /// Value shuffle:
2349 ///   In this step active threads transfer data from higher lane positions
2350 ///   in the warp to lower lane positions, creating Remote Reduce list.
2351 /// Value aggregation:
2352 ///   In this step, effective threads combine their thread local Reduce list
2353 ///   with Remote Reduce list and store the result in the thread local
2354 ///   Reduce list.
2355 /// Value copy:
2356 ///   In this step, we deal with the assumption made by algorithm 2
2357 ///   (i.e. contiguity assumption).  When we have an odd number of lanes
2358 ///   active, say 2k+1, only k threads will be effective and therefore k
2359 ///   new values will be produced.  However, the Reduce list owned by the
2360 ///   (2k+1)th thread is ignored in the value aggregation.  Therefore
2361 ///   we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so
2362 ///   that the contiguity assumption still holds.
2363 static llvm::Function *emitShuffleAndReduceFunction(
2364     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2365     QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) {
2366   ASTContext &C = CGM.getContext();
2367 
2368   // Thread local Reduce list used to host the values of data to be reduced.
2369   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2370                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2371   // Current lane id; could be logical.
2372   ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy,
2373                               ImplicitParamDecl::Other);
2374   // Offset of the remote source lane relative to the current lane.
2375   ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2376                                         C.ShortTy, ImplicitParamDecl::Other);
2377   // Algorithm version.  This is expected to be known at compile time.
2378   ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2379                                C.ShortTy, ImplicitParamDecl::Other);
2380   FunctionArgList Args;
2381   Args.push_back(&ReduceListArg);
2382   Args.push_back(&LaneIDArg);
2383   Args.push_back(&RemoteLaneOffsetArg);
2384   Args.push_back(&AlgoVerArg);
2385 
2386   const CGFunctionInfo &CGFI =
2387       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2388   auto *Fn = llvm::Function::Create(
2389       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2390       "_omp_reduction_shuffle_and_reduce_func", &CGM.getModule());
2391   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2392   Fn->setDoesNotRecurse();
2393 
2394   CodeGenFunction CGF(CGM);
2395   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2396 
2397   CGBuilderTy &Bld = CGF.Builder;
2398 
2399   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2400   Address LocalReduceList(
2401       Bld.CreatePointerBitCastOrAddrSpaceCast(
2402           CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
2403                                C.VoidPtrTy, SourceLocation()),
2404           CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
2405       CGF.getPointerAlign());
2406 
2407   Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg);
2408   llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar(
2409       AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
2410 
2411   Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg);
2412   llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar(
2413       AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
2414 
2415   Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg);
2416   llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar(
2417       AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
2418 
2419   // Create a local thread-private variable to host the Reduce list
2420   // from a remote lane.
2421   Address RemoteReduceList =
2422       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list");
2423 
2424   // This loop iterates through the list of reduce elements and copies,
2425   // element by element, from a remote lane in the warp to RemoteReduceList,
2426   // hosted on the thread's stack.
2427   emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates,
2428                         LocalReduceList, RemoteReduceList,
2429                         {/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal,
2430                          /*ScratchpadIndex=*/nullptr,
2431                          /*ScratchpadWidth=*/nullptr});
2432 
2433   // The actions to be performed on the Remote Reduce list is dependent
2434   // on the algorithm version.
2435   //
2436   //  if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 &&
2437   //  LaneId % 2 == 0 && Offset > 0):
2438   //    do the reduction value aggregation
2439   //
2440   //  The thread local variable Reduce list is mutated in place to host the
2441   //  reduced data, which is the aggregated value produced from local and
2442   //  remote lanes.
2443   //
2444   //  Note that AlgoVer is expected to be a constant integer known at compile
2445   //  time.
2446   //  When AlgoVer==0, the first conjunction evaluates to true, making
2447   //    the entire predicate true during compile time.
2448   //  When AlgoVer==1, the second conjunction has only the second part to be
2449   //    evaluated during runtime.  Other conjunctions evaluates to false
2450   //    during compile time.
2451   //  When AlgoVer==2, the third conjunction has only the second part to be
2452   //    evaluated during runtime.  Other conjunctions evaluates to false
2453   //    during compile time.
2454   llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal);
2455 
2456   llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1));
2457   llvm::Value *CondAlgo1 = Bld.CreateAnd(
2458       Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal));
2459 
2460   llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2));
2461   llvm::Value *CondAlgo2 = Bld.CreateAnd(
2462       Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1))));
2463   CondAlgo2 = Bld.CreateAnd(
2464       CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0)));
2465 
2466   llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1);
2467   CondReduce = Bld.CreateOr(CondReduce, CondAlgo2);
2468 
2469   llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then");
2470   llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else");
2471   llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont");
2472   Bld.CreateCondBr(CondReduce, ThenBB, ElseBB);
2473 
2474   CGF.EmitBlock(ThenBB);
2475   // reduce_function(LocalReduceList, RemoteReduceList)
2476   llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2477       LocalReduceList.getPointer(), CGF.VoidPtrTy);
2478   llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2479       RemoteReduceList.getPointer(), CGF.VoidPtrTy);
2480   CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
2481       CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr});
2482   Bld.CreateBr(MergeBB);
2483 
2484   CGF.EmitBlock(ElseBB);
2485   Bld.CreateBr(MergeBB);
2486 
2487   CGF.EmitBlock(MergeBB);
2488 
2489   // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local
2490   // Reduce list.
2491   Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1));
2492   llvm::Value *CondCopy = Bld.CreateAnd(
2493       Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal));
2494 
2495   llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then");
2496   llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else");
2497   llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont");
2498   Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB);
2499 
2500   CGF.EmitBlock(CpyThenBB);
2501   emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates,
2502                         RemoteReduceList, LocalReduceList);
2503   Bld.CreateBr(CpyMergeBB);
2504 
2505   CGF.EmitBlock(CpyElseBB);
2506   Bld.CreateBr(CpyMergeBB);
2507 
2508   CGF.EmitBlock(CpyMergeBB);
2509 
2510   CGF.FinishFunction();
2511   return Fn;
2512 }
2513 
2514 /// This function emits a helper that copies all the reduction variables from
2515 /// the team into the provided global buffer for the reduction variables.
2516 ///
2517 /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
2518 ///   For all data entries D in reduce_data:
2519 ///     Copy local D to buffer.D[Idx]
2520 static llvm::Value *emitListToGlobalCopyFunction(
2521     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2522     QualType ReductionArrayTy, SourceLocation Loc,
2523     const RecordDecl *TeamReductionRec,
2524     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2525         &VarFieldMap) {
2526   ASTContext &C = CGM.getContext();
2527 
2528   // Buffer: global reduction buffer.
2529   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2530                               C.VoidPtrTy, ImplicitParamDecl::Other);
2531   // Idx: index of the buffer.
2532   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2533                            ImplicitParamDecl::Other);
2534   // ReduceList: thread local Reduce list.
2535   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2536                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2537   FunctionArgList Args;
2538   Args.push_back(&BufferArg);
2539   Args.push_back(&IdxArg);
2540   Args.push_back(&ReduceListArg);
2541 
2542   const CGFunctionInfo &CGFI =
2543       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2544   auto *Fn = llvm::Function::Create(
2545       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2546       "_omp_reduction_list_to_global_copy_func", &CGM.getModule());
2547   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2548   Fn->setDoesNotRecurse();
2549   CodeGenFunction CGF(CGM);
2550   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2551 
2552   CGBuilderTy &Bld = CGF.Builder;
2553 
2554   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2555   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2556   Address LocalReduceList(
2557       Bld.CreatePointerBitCastOrAddrSpaceCast(
2558           CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
2559                                C.VoidPtrTy, Loc),
2560           CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
2561       CGF.getPointerAlign());
2562   QualType StaticTy = C.getRecordType(TeamReductionRec);
2563   llvm::Type *LLVMReductionsBufferTy =
2564       CGM.getTypes().ConvertTypeForMem(StaticTy);
2565   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2566       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2567       LLVMReductionsBufferTy->getPointerTo());
2568   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2569                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2570                                               /*Volatile=*/false, C.IntTy,
2571                                               Loc)};
2572   unsigned Idx = 0;
2573   for (const Expr *Private : Privates) {
2574     // Reduce element = LocalReduceList[i]
2575     Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
2576     llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
2577         ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
2578     // elemptr = ((CopyType*)(elemptrptr)) + I
2579     ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2580         ElemPtrPtr, CGF.ConvertTypeForMem(Private->getType())->getPointerTo());
2581     Address ElemPtr =
2582         Address(ElemPtrPtr, C.getTypeAlignInChars(Private->getType()));
2583     const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl();
2584     // Global = Buffer.VD[Idx];
2585     const FieldDecl *FD = VarFieldMap.lookup(VD);
2586     LValue GlobLVal = CGF.EmitLValueForField(
2587         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2588     Address GlobAddr = GlobLVal.getAddress(CGF);
2589     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(
2590         GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs);
2591     GlobLVal.setAddress(Address(BufferPtr, GlobAddr.getAlignment()));
2592     switch (CGF.getEvaluationKind(Private->getType())) {
2593     case TEK_Scalar: {
2594       llvm::Value *V = CGF.EmitLoadOfScalar(
2595           ElemPtr, /*Volatile=*/false, Private->getType(), Loc,
2596           LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
2597       CGF.EmitStoreOfScalar(V, GlobLVal);
2598       break;
2599     }
2600     case TEK_Complex: {
2601       CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(
2602           CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc);
2603       CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false);
2604       break;
2605     }
2606     case TEK_Aggregate:
2607       CGF.EmitAggregateCopy(GlobLVal,
2608                             CGF.MakeAddrLValue(ElemPtr, Private->getType()),
2609                             Private->getType(), AggValueSlot::DoesNotOverlap);
2610       break;
2611     }
2612     ++Idx;
2613   }
2614 
2615   CGF.FinishFunction();
2616   return Fn;
2617 }
2618 
2619 /// This function emits a helper that reduces all the reduction variables from
2620 /// the team into the provided global buffer for the reduction variables.
2621 ///
2622 /// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data)
2623 ///  void *GlobPtrs[];
2624 ///  GlobPtrs[0] = (void*)&buffer.D0[Idx];
2625 ///  ...
2626 ///  GlobPtrs[N] = (void*)&buffer.DN[Idx];
2627 ///  reduce_function(GlobPtrs, reduce_data);
2628 static llvm::Value *emitListToGlobalReduceFunction(
2629     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2630     QualType ReductionArrayTy, SourceLocation Loc,
2631     const RecordDecl *TeamReductionRec,
2632     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2633         &VarFieldMap,
2634     llvm::Function *ReduceFn) {
2635   ASTContext &C = CGM.getContext();
2636 
2637   // Buffer: global reduction buffer.
2638   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2639                               C.VoidPtrTy, ImplicitParamDecl::Other);
2640   // Idx: index of the buffer.
2641   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2642                            ImplicitParamDecl::Other);
2643   // ReduceList: thread local Reduce list.
2644   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2645                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2646   FunctionArgList Args;
2647   Args.push_back(&BufferArg);
2648   Args.push_back(&IdxArg);
2649   Args.push_back(&ReduceListArg);
2650 
2651   const CGFunctionInfo &CGFI =
2652       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2653   auto *Fn = llvm::Function::Create(
2654       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2655       "_omp_reduction_list_to_global_reduce_func", &CGM.getModule());
2656   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2657   Fn->setDoesNotRecurse();
2658   CodeGenFunction CGF(CGM);
2659   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2660 
2661   CGBuilderTy &Bld = CGF.Builder;
2662 
2663   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2664   QualType StaticTy = C.getRecordType(TeamReductionRec);
2665   llvm::Type *LLVMReductionsBufferTy =
2666       CGM.getTypes().ConvertTypeForMem(StaticTy);
2667   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2668       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2669       LLVMReductionsBufferTy->getPointerTo());
2670 
2671   // 1. Build a list of reduction variables.
2672   // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
2673   Address ReductionList =
2674       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
2675   auto IPriv = Privates.begin();
2676   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2677                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2678                                               /*Volatile=*/false, C.IntTy,
2679                                               Loc)};
2680   unsigned Idx = 0;
2681   for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) {
2682     Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2683     // Global = Buffer.VD[Idx];
2684     const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl();
2685     const FieldDecl *FD = VarFieldMap.lookup(VD);
2686     LValue GlobLVal = CGF.EmitLValueForField(
2687         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2688     Address GlobAddr = GlobLVal.getAddress(CGF);
2689     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(
2690         GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs);
2691     llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr);
2692     CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy);
2693     if ((*IPriv)->getType()->isVariablyModifiedType()) {
2694       // Store array size.
2695       ++Idx;
2696       Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2697       llvm::Value *Size = CGF.Builder.CreateIntCast(
2698           CGF.getVLASize(
2699                  CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
2700               .NumElts,
2701           CGF.SizeTy, /*isSigned=*/false);
2702       CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
2703                               Elem);
2704     }
2705   }
2706 
2707   // Call reduce_function(GlobalReduceList, ReduceList)
2708   llvm::Value *GlobalReduceList =
2709       CGF.EmitCastToVoidPtr(ReductionList.getPointer());
2710   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2711   llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar(
2712       AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc);
2713   CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
2714       CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr});
2715   CGF.FinishFunction();
2716   return Fn;
2717 }
2718 
2719 /// This function emits a helper that copies all the reduction variables from
2720 /// the team into the provided global buffer for the reduction variables.
2721 ///
2722 /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
2723 ///   For all data entries D in reduce_data:
2724 ///     Copy buffer.D[Idx] to local D;
2725 static llvm::Value *emitGlobalToListCopyFunction(
2726     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2727     QualType ReductionArrayTy, SourceLocation Loc,
2728     const RecordDecl *TeamReductionRec,
2729     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2730         &VarFieldMap) {
2731   ASTContext &C = CGM.getContext();
2732 
2733   // Buffer: global reduction buffer.
2734   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2735                               C.VoidPtrTy, ImplicitParamDecl::Other);
2736   // Idx: index of the buffer.
2737   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2738                            ImplicitParamDecl::Other);
2739   // ReduceList: thread local Reduce list.
2740   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2741                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2742   FunctionArgList Args;
2743   Args.push_back(&BufferArg);
2744   Args.push_back(&IdxArg);
2745   Args.push_back(&ReduceListArg);
2746 
2747   const CGFunctionInfo &CGFI =
2748       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2749   auto *Fn = llvm::Function::Create(
2750       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2751       "_omp_reduction_global_to_list_copy_func", &CGM.getModule());
2752   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2753   Fn->setDoesNotRecurse();
2754   CodeGenFunction CGF(CGM);
2755   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2756 
2757   CGBuilderTy &Bld = CGF.Builder;
2758 
2759   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2760   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2761   Address LocalReduceList(
2762       Bld.CreatePointerBitCastOrAddrSpaceCast(
2763           CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
2764                                C.VoidPtrTy, Loc),
2765           CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo()),
2766       CGF.getPointerAlign());
2767   QualType StaticTy = C.getRecordType(TeamReductionRec);
2768   llvm::Type *LLVMReductionsBufferTy =
2769       CGM.getTypes().ConvertTypeForMem(StaticTy);
2770   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2771       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2772       LLVMReductionsBufferTy->getPointerTo());
2773 
2774   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2775                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2776                                               /*Volatile=*/false, C.IntTy,
2777                                               Loc)};
2778   unsigned Idx = 0;
2779   for (const Expr *Private : Privates) {
2780     // Reduce element = LocalReduceList[i]
2781     Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
2782     llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
2783         ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
2784     // elemptr = ((CopyType*)(elemptrptr)) + I
2785     ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2786         ElemPtrPtr, CGF.ConvertTypeForMem(Private->getType())->getPointerTo());
2787     Address ElemPtr =
2788         Address(ElemPtrPtr, C.getTypeAlignInChars(Private->getType()));
2789     const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl();
2790     // Global = Buffer.VD[Idx];
2791     const FieldDecl *FD = VarFieldMap.lookup(VD);
2792     LValue GlobLVal = CGF.EmitLValueForField(
2793         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2794     Address GlobAddr = GlobLVal.getAddress(CGF);
2795     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(
2796         GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs);
2797     GlobLVal.setAddress(Address(BufferPtr, GlobAddr.getAlignment()));
2798     switch (CGF.getEvaluationKind(Private->getType())) {
2799     case TEK_Scalar: {
2800       llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc);
2801       CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType(),
2802                             LValueBaseInfo(AlignmentSource::Type),
2803                             TBAAAccessInfo());
2804       break;
2805     }
2806     case TEK_Complex: {
2807       CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc);
2808       CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()),
2809                              /*isInit=*/false);
2810       break;
2811     }
2812     case TEK_Aggregate:
2813       CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()),
2814                             GlobLVal, Private->getType(),
2815                             AggValueSlot::DoesNotOverlap);
2816       break;
2817     }
2818     ++Idx;
2819   }
2820 
2821   CGF.FinishFunction();
2822   return Fn;
2823 }
2824 
2825 /// This function emits a helper that reduces all the reduction variables from
2826 /// the team into the provided global buffer for the reduction variables.
2827 ///
2828 /// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data)
2829 ///  void *GlobPtrs[];
2830 ///  GlobPtrs[0] = (void*)&buffer.D0[Idx];
2831 ///  ...
2832 ///  GlobPtrs[N] = (void*)&buffer.DN[Idx];
2833 ///  reduce_function(reduce_data, GlobPtrs);
2834 static llvm::Value *emitGlobalToListReduceFunction(
2835     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2836     QualType ReductionArrayTy, SourceLocation Loc,
2837     const RecordDecl *TeamReductionRec,
2838     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2839         &VarFieldMap,
2840     llvm::Function *ReduceFn) {
2841   ASTContext &C = CGM.getContext();
2842 
2843   // Buffer: global reduction buffer.
2844   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2845                               C.VoidPtrTy, ImplicitParamDecl::Other);
2846   // Idx: index of the buffer.
2847   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2848                            ImplicitParamDecl::Other);
2849   // ReduceList: thread local Reduce list.
2850   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2851                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2852   FunctionArgList Args;
2853   Args.push_back(&BufferArg);
2854   Args.push_back(&IdxArg);
2855   Args.push_back(&ReduceListArg);
2856 
2857   const CGFunctionInfo &CGFI =
2858       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2859   auto *Fn = llvm::Function::Create(
2860       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2861       "_omp_reduction_global_to_list_reduce_func", &CGM.getModule());
2862   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2863   Fn->setDoesNotRecurse();
2864   CodeGenFunction CGF(CGM);
2865   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2866 
2867   CGBuilderTy &Bld = CGF.Builder;
2868 
2869   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2870   QualType StaticTy = C.getRecordType(TeamReductionRec);
2871   llvm::Type *LLVMReductionsBufferTy =
2872       CGM.getTypes().ConvertTypeForMem(StaticTy);
2873   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2874       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2875       LLVMReductionsBufferTy->getPointerTo());
2876 
2877   // 1. Build a list of reduction variables.
2878   // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
2879   Address ReductionList =
2880       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
2881   auto IPriv = Privates.begin();
2882   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2883                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2884                                               /*Volatile=*/false, C.IntTy,
2885                                               Loc)};
2886   unsigned Idx = 0;
2887   for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) {
2888     Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2889     // Global = Buffer.VD[Idx];
2890     const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl();
2891     const FieldDecl *FD = VarFieldMap.lookup(VD);
2892     LValue GlobLVal = CGF.EmitLValueForField(
2893         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2894     Address GlobAddr = GlobLVal.getAddress(CGF);
2895     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(
2896         GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs);
2897     llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr);
2898     CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy);
2899     if ((*IPriv)->getType()->isVariablyModifiedType()) {
2900       // Store array size.
2901       ++Idx;
2902       Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2903       llvm::Value *Size = CGF.Builder.CreateIntCast(
2904           CGF.getVLASize(
2905                  CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
2906               .NumElts,
2907           CGF.SizeTy, /*isSigned=*/false);
2908       CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
2909                               Elem);
2910     }
2911   }
2912 
2913   // Call reduce_function(ReduceList, GlobalReduceList)
2914   llvm::Value *GlobalReduceList =
2915       CGF.EmitCastToVoidPtr(ReductionList.getPointer());
2916   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2917   llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar(
2918       AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc);
2919   CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
2920       CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList});
2921   CGF.FinishFunction();
2922   return Fn;
2923 }
2924 
2925 ///
2926 /// Design of OpenMP reductions on the GPU
2927 ///
2928 /// Consider a typical OpenMP program with one or more reduction
2929 /// clauses:
2930 ///
2931 /// float foo;
2932 /// double bar;
2933 /// #pragma omp target teams distribute parallel for \
2934 ///             reduction(+:foo) reduction(*:bar)
2935 /// for (int i = 0; i < N; i++) {
2936 ///   foo += A[i]; bar *= B[i];
2937 /// }
2938 ///
2939 /// where 'foo' and 'bar' are reduced across all OpenMP threads in
2940 /// all teams.  In our OpenMP implementation on the NVPTX device an
2941 /// OpenMP team is mapped to a CUDA threadblock and OpenMP threads
2942 /// within a team are mapped to CUDA threads within a threadblock.
2943 /// Our goal is to efficiently aggregate values across all OpenMP
2944 /// threads such that:
2945 ///
2946 ///   - the compiler and runtime are logically concise, and
2947 ///   - the reduction is performed efficiently in a hierarchical
2948 ///     manner as follows: within OpenMP threads in the same warp,
2949 ///     across warps in a threadblock, and finally across teams on
2950 ///     the NVPTX device.
2951 ///
2952 /// Introduction to Decoupling
2953 ///
2954 /// We would like to decouple the compiler and the runtime so that the
2955 /// latter is ignorant of the reduction variables (number, data types)
2956 /// and the reduction operators.  This allows a simpler interface
2957 /// and implementation while still attaining good performance.
2958 ///
2959 /// Pseudocode for the aforementioned OpenMP program generated by the
2960 /// compiler is as follows:
2961 ///
2962 /// 1. Create private copies of reduction variables on each OpenMP
2963 ///    thread: 'foo_private', 'bar_private'
2964 /// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned
2965 ///    to it and writes the result in 'foo_private' and 'bar_private'
2966 ///    respectively.
2967 /// 3. Call the OpenMP runtime on the GPU to reduce within a team
2968 ///    and store the result on the team master:
2969 ///
2970 ///     __kmpc_nvptx_parallel_reduce_nowait_v2(...,
2971 ///        reduceData, shuffleReduceFn, interWarpCpyFn)
2972 ///
2973 ///     where:
2974 ///       struct ReduceData {
2975 ///         double *foo;
2976 ///         double *bar;
2977 ///       } reduceData
2978 ///       reduceData.foo = &foo_private
2979 ///       reduceData.bar = &bar_private
2980 ///
2981 ///     'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two
2982 ///     auxiliary functions generated by the compiler that operate on
2983 ///     variables of type 'ReduceData'.  They aid the runtime perform
2984 ///     algorithmic steps in a data agnostic manner.
2985 ///
2986 ///     'shuffleReduceFn' is a pointer to a function that reduces data
2987 ///     of type 'ReduceData' across two OpenMP threads (lanes) in the
2988 ///     same warp.  It takes the following arguments as input:
2989 ///
2990 ///     a. variable of type 'ReduceData' on the calling lane,
2991 ///     b. its lane_id,
2992 ///     c. an offset relative to the current lane_id to generate a
2993 ///        remote_lane_id.  The remote lane contains the second
2994 ///        variable of type 'ReduceData' that is to be reduced.
2995 ///     d. an algorithm version parameter determining which reduction
2996 ///        algorithm to use.
2997 ///
2998 ///     'shuffleReduceFn' retrieves data from the remote lane using
2999 ///     efficient GPU shuffle intrinsics and reduces, using the
3000 ///     algorithm specified by the 4th parameter, the two operands
3001 ///     element-wise.  The result is written to the first operand.
3002 ///
3003 ///     Different reduction algorithms are implemented in different
3004 ///     runtime functions, all calling 'shuffleReduceFn' to perform
3005 ///     the essential reduction step.  Therefore, based on the 4th
3006 ///     parameter, this function behaves slightly differently to
3007 ///     cooperate with the runtime to ensure correctness under
3008 ///     different circumstances.
3009 ///
3010 ///     'InterWarpCpyFn' is a pointer to a function that transfers
3011 ///     reduced variables across warps.  It tunnels, through CUDA
3012 ///     shared memory, the thread-private data of type 'ReduceData'
3013 ///     from lane 0 of each warp to a lane in the first warp.
3014 /// 4. Call the OpenMP runtime on the GPU to reduce across teams.
3015 ///    The last team writes the global reduced value to memory.
3016 ///
3017 ///     ret = __kmpc_nvptx_teams_reduce_nowait(...,
3018 ///             reduceData, shuffleReduceFn, interWarpCpyFn,
3019 ///             scratchpadCopyFn, loadAndReduceFn)
3020 ///
3021 ///     'scratchpadCopyFn' is a helper that stores reduced
3022 ///     data from the team master to a scratchpad array in
3023 ///     global memory.
3024 ///
3025 ///     'loadAndReduceFn' is a helper that loads data from
3026 ///     the scratchpad array and reduces it with the input
3027 ///     operand.
3028 ///
3029 ///     These compiler generated functions hide address
3030 ///     calculation and alignment information from the runtime.
3031 /// 5. if ret == 1:
3032 ///     The team master of the last team stores the reduced
3033 ///     result to the globals in memory.
3034 ///     foo += reduceData.foo; bar *= reduceData.bar
3035 ///
3036 ///
3037 /// Warp Reduction Algorithms
3038 ///
3039 /// On the warp level, we have three algorithms implemented in the
3040 /// OpenMP runtime depending on the number of active lanes:
3041 ///
3042 /// Full Warp Reduction
3043 ///
3044 /// The reduce algorithm within a warp where all lanes are active
3045 /// is implemented in the runtime as follows:
3046 ///
3047 /// full_warp_reduce(void *reduce_data,
3048 ///                  kmp_ShuffleReductFctPtr ShuffleReduceFn) {
3049 ///   for (int offset = WARPSIZE/2; offset > 0; offset /= 2)
3050 ///     ShuffleReduceFn(reduce_data, 0, offset, 0);
3051 /// }
3052 ///
3053 /// The algorithm completes in log(2, WARPSIZE) steps.
3054 ///
3055 /// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is
3056 /// not used therefore we save instructions by not retrieving lane_id
3057 /// from the corresponding special registers.  The 4th parameter, which
3058 /// represents the version of the algorithm being used, is set to 0 to
3059 /// signify full warp reduction.
3060 ///
3061 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
3062 ///
3063 /// #reduce_elem refers to an element in the local lane's data structure
3064 /// #remote_elem is retrieved from a remote lane
3065 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
3066 /// reduce_elem = reduce_elem REDUCE_OP remote_elem;
3067 ///
3068 /// Contiguous Partial Warp Reduction
3069 ///
3070 /// This reduce algorithm is used within a warp where only the first
3071 /// 'n' (n <= WARPSIZE) lanes are active.  It is typically used when the
3072 /// number of OpenMP threads in a parallel region is not a multiple of
3073 /// WARPSIZE.  The algorithm is implemented in the runtime as follows:
3074 ///
3075 /// void
3076 /// contiguous_partial_reduce(void *reduce_data,
3077 ///                           kmp_ShuffleReductFctPtr ShuffleReduceFn,
3078 ///                           int size, int lane_id) {
3079 ///   int curr_size;
3080 ///   int offset;
3081 ///   curr_size = size;
3082 ///   mask = curr_size/2;
3083 ///   while (offset>0) {
3084 ///     ShuffleReduceFn(reduce_data, lane_id, offset, 1);
3085 ///     curr_size = (curr_size+1)/2;
3086 ///     offset = curr_size/2;
3087 ///   }
3088 /// }
3089 ///
3090 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
3091 ///
3092 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
3093 /// if (lane_id < offset)
3094 ///     reduce_elem = reduce_elem REDUCE_OP remote_elem
3095 /// else
3096 ///     reduce_elem = remote_elem
3097 ///
3098 /// This algorithm assumes that the data to be reduced are located in a
3099 /// contiguous subset of lanes starting from the first.  When there is
3100 /// an odd number of active lanes, the data in the last lane is not
3101 /// aggregated with any other lane's dat but is instead copied over.
3102 ///
3103 /// Dispersed Partial Warp Reduction
3104 ///
3105 /// This algorithm is used within a warp when any discontiguous subset of
3106 /// lanes are active.  It is used to implement the reduction operation
3107 /// across lanes in an OpenMP simd region or in a nested parallel region.
3108 ///
3109 /// void
3110 /// dispersed_partial_reduce(void *reduce_data,
3111 ///                          kmp_ShuffleReductFctPtr ShuffleReduceFn) {
3112 ///   int size, remote_id;
3113 ///   int logical_lane_id = number_of_active_lanes_before_me() * 2;
3114 ///   do {
3115 ///       remote_id = next_active_lane_id_right_after_me();
3116 ///       # the above function returns 0 of no active lane
3117 ///       # is present right after the current lane.
3118 ///       size = number_of_active_lanes_in_this_warp();
3119 ///       logical_lane_id /= 2;
3120 ///       ShuffleReduceFn(reduce_data, logical_lane_id,
3121 ///                       remote_id-1-threadIdx.x, 2);
3122 ///   } while (logical_lane_id % 2 == 0 && size > 1);
3123 /// }
3124 ///
3125 /// There is no assumption made about the initial state of the reduction.
3126 /// Any number of lanes (>=1) could be active at any position.  The reduction
3127 /// result is returned in the first active lane.
3128 ///
3129 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
3130 ///
3131 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
3132 /// if (lane_id % 2 == 0 && offset > 0)
3133 ///     reduce_elem = reduce_elem REDUCE_OP remote_elem
3134 /// else
3135 ///     reduce_elem = remote_elem
3136 ///
3137 ///
3138 /// Intra-Team Reduction
3139 ///
3140 /// This function, as implemented in the runtime call
3141 /// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP
3142 /// threads in a team.  It first reduces within a warp using the
3143 /// aforementioned algorithms.  We then proceed to gather all such
3144 /// reduced values at the first warp.
3145 ///
3146 /// The runtime makes use of the function 'InterWarpCpyFn', which copies
3147 /// data from each of the "warp master" (zeroth lane of each warp, where
3148 /// warp-reduced data is held) to the zeroth warp.  This step reduces (in
3149 /// a mathematical sense) the problem of reduction across warp masters in
3150 /// a block to the problem of warp reduction.
3151 ///
3152 ///
3153 /// Inter-Team Reduction
3154 ///
3155 /// Once a team has reduced its data to a single value, it is stored in
3156 /// a global scratchpad array.  Since each team has a distinct slot, this
3157 /// can be done without locking.
3158 ///
3159 /// The last team to write to the scratchpad array proceeds to reduce the
3160 /// scratchpad array.  One or more workers in the last team use the helper
3161 /// 'loadAndReduceDataFn' to load and reduce values from the array, i.e.,
3162 /// the k'th worker reduces every k'th element.
3163 ///
3164 /// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to
3165 /// reduce across workers and compute a globally reduced value.
3166 ///
3167 void CGOpenMPRuntimeGPU::emitReduction(
3168     CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates,
3169     ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs,
3170     ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) {
3171   if (!CGF.HaveInsertPoint())
3172     return;
3173 
3174   bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind);
3175 #ifndef NDEBUG
3176   bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind);
3177 #endif
3178 
3179   if (Options.SimpleReduction) {
3180     assert(!TeamsReduction && !ParallelReduction &&
3181            "Invalid reduction selection in emitReduction.");
3182     CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs,
3183                                    ReductionOps, Options);
3184     return;
3185   }
3186 
3187   assert((TeamsReduction || ParallelReduction) &&
3188          "Invalid reduction selection in emitReduction.");
3189 
3190   // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList),
3191   // RedList, shuffle_reduce_func, interwarp_copy_func);
3192   // or
3193   // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>);
3194   llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
3195   llvm::Value *ThreadId = getThreadID(CGF, Loc);
3196 
3197   llvm::Value *Res;
3198   ASTContext &C = CGM.getContext();
3199   // 1. Build a list of reduction variables.
3200   // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
3201   auto Size = RHSExprs.size();
3202   for (const Expr *E : Privates) {
3203     if (E->getType()->isVariablyModifiedType())
3204       // Reserve place for array size.
3205       ++Size;
3206   }
3207   llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size);
3208   QualType ReductionArrayTy =
3209       C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal,
3210                              /*IndexTypeQuals=*/0);
3211   Address ReductionList =
3212       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
3213   auto IPriv = Privates.begin();
3214   unsigned Idx = 0;
3215   for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) {
3216     Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
3217     CGF.Builder.CreateStore(
3218         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3219             CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy),
3220         Elem);
3221     if ((*IPriv)->getType()->isVariablyModifiedType()) {
3222       // Store array size.
3223       ++Idx;
3224       Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
3225       llvm::Value *Size = CGF.Builder.CreateIntCast(
3226           CGF.getVLASize(
3227                  CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
3228               .NumElts,
3229           CGF.SizeTy, /*isSigned=*/false);
3230       CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
3231                               Elem);
3232     }
3233   }
3234 
3235   llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3236       ReductionList.getPointer(), CGF.VoidPtrTy);
3237   llvm::Function *ReductionFn = emitReductionFunction(
3238       Loc, CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo(), Privates,
3239       LHSExprs, RHSExprs, ReductionOps);
3240   llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy);
3241   llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction(
3242       CGM, Privates, ReductionArrayTy, ReductionFn, Loc);
3243   llvm::Value *InterWarpCopyFn =
3244       emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc);
3245 
3246   if (ParallelReduction) {
3247     llvm::Value *Args[] = {RTLoc,
3248                            ThreadId,
3249                            CGF.Builder.getInt32(RHSExprs.size()),
3250                            ReductionArrayTySize,
3251                            RL,
3252                            ShuffleAndReduceFn,
3253                            InterWarpCopyFn};
3254 
3255     Res = CGF.EmitRuntimeCall(
3256         OMPBuilder.getOrCreateRuntimeFunction(
3257             CGM.getModule(), OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2),
3258         Args);
3259   } else {
3260     assert(TeamsReduction && "expected teams reduction.");
3261     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap;
3262     llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size());
3263     int Cnt = 0;
3264     for (const Expr *DRE : Privates) {
3265       PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl();
3266       ++Cnt;
3267     }
3268     const RecordDecl *TeamReductionRec = ::buildRecordForGlobalizedVars(
3269         CGM.getContext(), PrivatesReductions, llvm::None, VarFieldMap,
3270         C.getLangOpts().OpenMPCUDAReductionBufNum);
3271     TeamsReductions.push_back(TeamReductionRec);
3272     if (!KernelTeamsReductionPtr) {
3273       KernelTeamsReductionPtr = new llvm::GlobalVariable(
3274           CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/true,
3275           llvm::GlobalValue::InternalLinkage, nullptr,
3276           "_openmp_teams_reductions_buffer_$_$ptr");
3277     }
3278     llvm::Value *GlobalBufferPtr = CGF.EmitLoadOfScalar(
3279         Address(KernelTeamsReductionPtr, CGM.getPointerAlign()),
3280         /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc);
3281     llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction(
3282         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap);
3283     llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction(
3284         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap,
3285         ReductionFn);
3286     llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction(
3287         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap);
3288     llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction(
3289         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap,
3290         ReductionFn);
3291 
3292     llvm::Value *Args[] = {
3293         RTLoc,
3294         ThreadId,
3295         GlobalBufferPtr,
3296         CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum),
3297         RL,
3298         ShuffleAndReduceFn,
3299         InterWarpCopyFn,
3300         GlobalToBufferCpyFn,
3301         GlobalToBufferRedFn,
3302         BufferToGlobalCpyFn,
3303         BufferToGlobalRedFn};
3304 
3305     Res = CGF.EmitRuntimeCall(
3306         OMPBuilder.getOrCreateRuntimeFunction(
3307             CGM.getModule(), OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2),
3308         Args);
3309   }
3310 
3311   // 5. Build if (res == 1)
3312   llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done");
3313   llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then");
3314   llvm::Value *Cond = CGF.Builder.CreateICmpEQ(
3315       Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1));
3316   CGF.Builder.CreateCondBr(Cond, ThenBB, ExitBB);
3317 
3318   // 6. Build then branch: where we have reduced values in the master
3319   //    thread in each team.
3320   //    __kmpc_end_reduce{_nowait}(<gtid>);
3321   //    break;
3322   CGF.EmitBlock(ThenBB);
3323 
3324   // Add emission of __kmpc_end_reduce{_nowait}(<gtid>);
3325   auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps,
3326                     this](CodeGenFunction &CGF, PrePostActionTy &Action) {
3327     auto IPriv = Privates.begin();
3328     auto ILHS = LHSExprs.begin();
3329     auto IRHS = RHSExprs.begin();
3330     for (const Expr *E : ReductionOps) {
3331       emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
3332                                   cast<DeclRefExpr>(*IRHS));
3333       ++IPriv;
3334       ++ILHS;
3335       ++IRHS;
3336     }
3337   };
3338   llvm::Value *EndArgs[] = {ThreadId};
3339   RegionCodeGenTy RCG(CodeGen);
3340   NVPTXActionTy Action(
3341       nullptr, llvm::None,
3342       OMPBuilder.getOrCreateRuntimeFunction(
3343           CGM.getModule(), OMPRTL___kmpc_nvptx_end_reduce_nowait),
3344       EndArgs);
3345   RCG.setAction(Action);
3346   RCG(CGF);
3347   // There is no need to emit line number for unconditional branch.
3348   (void)ApplyDebugLocation::CreateEmpty(CGF);
3349   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
3350 }
3351 
3352 const VarDecl *
3353 CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD,
3354                                        const VarDecl *NativeParam) const {
3355   if (!NativeParam->getType()->isReferenceType())
3356     return NativeParam;
3357   QualType ArgType = NativeParam->getType();
3358   QualifierCollector QC;
3359   const Type *NonQualTy = QC.strip(ArgType);
3360   QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
3361   if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) {
3362     if (Attr->getCaptureKind() == OMPC_map) {
3363       PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy,
3364                                                         LangAS::opencl_global);
3365     }
3366   }
3367   ArgType = CGM.getContext().getPointerType(PointeeTy);
3368   QC.addRestrict();
3369   enum { NVPTX_local_addr = 5 };
3370   QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr));
3371   ArgType = QC.apply(CGM.getContext(), ArgType);
3372   if (isa<ImplicitParamDecl>(NativeParam))
3373     return ImplicitParamDecl::Create(
3374         CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(),
3375         NativeParam->getIdentifier(), ArgType, ImplicitParamDecl::Other);
3376   return ParmVarDecl::Create(
3377       CGM.getContext(),
3378       const_cast<DeclContext *>(NativeParam->getDeclContext()),
3379       NativeParam->getBeginLoc(), NativeParam->getLocation(),
3380       NativeParam->getIdentifier(), ArgType,
3381       /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr);
3382 }
3383 
3384 Address
3385 CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF,
3386                                           const VarDecl *NativeParam,
3387                                           const VarDecl *TargetParam) const {
3388   assert(NativeParam != TargetParam &&
3389          NativeParam->getType()->isReferenceType() &&
3390          "Native arg must not be the same as target arg.");
3391   Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam);
3392   QualType NativeParamType = NativeParam->getType();
3393   QualifierCollector QC;
3394   const Type *NonQualTy = QC.strip(NativeParamType);
3395   QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
3396   unsigned NativePointeeAddrSpace =
3397       CGF.getContext().getTargetAddressSpace(NativePointeeTy);
3398   QualType TargetTy = TargetParam->getType();
3399   llvm::Value *TargetAddr = CGF.EmitLoadOfScalar(
3400       LocalAddr, /*Volatile=*/false, TargetTy, SourceLocation());
3401   // First cast to generic.
3402   TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3403       TargetAddr, TargetAddr->getType()->getPointerElementType()->getPointerTo(
3404                       /*AddrSpace=*/0));
3405   // Cast from generic to native address space.
3406   TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3407       TargetAddr, TargetAddr->getType()->getPointerElementType()->getPointerTo(
3408                       NativePointeeAddrSpace));
3409   Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType);
3410   CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false,
3411                         NativeParamType);
3412   return NativeParamAddr;
3413 }
3414 
3415 void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall(
3416     CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn,
3417     ArrayRef<llvm::Value *> Args) const {
3418   SmallVector<llvm::Value *, 4> TargetArgs;
3419   TargetArgs.reserve(Args.size());
3420   auto *FnType = OutlinedFn.getFunctionType();
3421   for (unsigned I = 0, E = Args.size(); I < E; ++I) {
3422     if (FnType->isVarArg() && FnType->getNumParams() <= I) {
3423       TargetArgs.append(std::next(Args.begin(), I), Args.end());
3424       break;
3425     }
3426     llvm::Type *TargetType = FnType->getParamType(I);
3427     llvm::Value *NativeArg = Args[I];
3428     if (!TargetType->isPointerTy()) {
3429       TargetArgs.emplace_back(NativeArg);
3430       continue;
3431     }
3432     llvm::Value *TargetArg = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3433         NativeArg,
3434         NativeArg->getType()->getPointerElementType()->getPointerTo());
3435     TargetArgs.emplace_back(
3436         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType));
3437   }
3438   CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs);
3439 }
3440 
3441 /// Emit function which wraps the outline parallel region
3442 /// and controls the arguments which are passed to this function.
3443 /// The wrapper ensures that the outlined function is called
3444 /// with the correct arguments when data is shared.
3445 llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper(
3446     llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) {
3447   ASTContext &Ctx = CGM.getContext();
3448   const auto &CS = *D.getCapturedStmt(OMPD_parallel);
3449 
3450   // Create a function that takes as argument the source thread.
3451   FunctionArgList WrapperArgs;
3452   QualType Int16QTy =
3453       Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false);
3454   QualType Int32QTy =
3455       Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false);
3456   ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
3457                                      /*Id=*/nullptr, Int16QTy,
3458                                      ImplicitParamDecl::Other);
3459   ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
3460                                /*Id=*/nullptr, Int32QTy,
3461                                ImplicitParamDecl::Other);
3462   WrapperArgs.emplace_back(&ParallelLevelArg);
3463   WrapperArgs.emplace_back(&WrapperArg);
3464 
3465   const CGFunctionInfo &CGFI =
3466       CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs);
3467 
3468   auto *Fn = llvm::Function::Create(
3469       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
3470       Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule());
3471 
3472   // Ensure we do not inline the function. This is trivially true for the ones
3473   // passed to __kmpc_fork_call but the ones calles in serialized regions
3474   // could be inlined. This is not a perfect but it is closer to the invariant
3475   // we want, namely, every data environment starts with a new function.
3476   // TODO: We should pass the if condition to the runtime function and do the
3477   //       handling there. Much cleaner code.
3478   Fn->addFnAttr(llvm::Attribute::NoInline);
3479 
3480   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
3481   Fn->setLinkage(llvm::GlobalValue::InternalLinkage);
3482   Fn->setDoesNotRecurse();
3483 
3484   CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
3485   CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs,
3486                     D.getBeginLoc(), D.getBeginLoc());
3487 
3488   const auto *RD = CS.getCapturedRecordDecl();
3489   auto CurField = RD->field_begin();
3490 
3491   Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
3492                                                       /*Name=*/".zero.addr");
3493   CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr);
3494   // Get the array of arguments.
3495   SmallVector<llvm::Value *, 8> Args;
3496 
3497   Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer());
3498   Args.emplace_back(ZeroAddr.getPointer());
3499 
3500   CGBuilderTy &Bld = CGF.Builder;
3501   auto CI = CS.capture_begin();
3502 
3503   // Use global memory for data sharing.
3504   // Handle passing of global args to workers.
3505   Address GlobalArgs =
3506       CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args");
3507   llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer();
3508   llvm::Value *DataSharingArgs[] = {GlobalArgsPtr};
3509   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
3510                           CGM.getModule(), OMPRTL___kmpc_get_shared_variables),
3511                       DataSharingArgs);
3512 
3513   // Retrieve the shared variables from the list of references returned
3514   // by the runtime. Pass the variables to the outlined function.
3515   Address SharedArgListAddress = Address::invalid();
3516   if (CS.capture_size() > 0 ||
3517       isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
3518     SharedArgListAddress = CGF.EmitLoadOfPointer(
3519         GlobalArgs, CGF.getContext()
3520                         .getPointerType(CGF.getContext().getPointerType(
3521                             CGF.getContext().VoidPtrTy))
3522                         .castAs<PointerType>());
3523   }
3524   unsigned Idx = 0;
3525   if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
3526     Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
3527     Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
3528         Src, CGF.SizeTy->getPointerTo());
3529     llvm::Value *LB = CGF.EmitLoadOfScalar(
3530         TypedAddress,
3531         /*Volatile=*/false,
3532         CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
3533         cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc());
3534     Args.emplace_back(LB);
3535     ++Idx;
3536     Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
3537     TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
3538         Src, CGF.SizeTy->getPointerTo());
3539     llvm::Value *UB = CGF.EmitLoadOfScalar(
3540         TypedAddress,
3541         /*Volatile=*/false,
3542         CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
3543         cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc());
3544     Args.emplace_back(UB);
3545     ++Idx;
3546   }
3547   if (CS.capture_size() > 0) {
3548     ASTContext &CGFContext = CGF.getContext();
3549     for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) {
3550       QualType ElemTy = CurField->getType();
3551       Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx);
3552       Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
3553           Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)));
3554       llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress,
3555                                               /*Volatile=*/false,
3556                                               CGFContext.getPointerType(ElemTy),
3557                                               CI->getLocation());
3558       if (CI->capturesVariableByCopy() &&
3559           !CI->getCapturedVar()->getType()->isAnyPointerType()) {
3560         Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(),
3561                               CI->getLocation());
3562       }
3563       Args.emplace_back(Arg);
3564     }
3565   }
3566 
3567   emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args);
3568   CGF.FinishFunction();
3569   return Fn;
3570 }
3571 
3572 void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF,
3573                                               const Decl *D) {
3574   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic)
3575     return;
3576 
3577   assert(D && "Expected function or captured|block decl.");
3578   assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 &&
3579          "Function is registered already.");
3580   assert((!TeamAndReductions.first || TeamAndReductions.first == D) &&
3581          "Team is set but not processed.");
3582   const Stmt *Body = nullptr;
3583   bool NeedToDelayGlobalization = false;
3584   if (const auto *FD = dyn_cast<FunctionDecl>(D)) {
3585     Body = FD->getBody();
3586   } else if (const auto *BD = dyn_cast<BlockDecl>(D)) {
3587     Body = BD->getBody();
3588   } else if (const auto *CD = dyn_cast<CapturedDecl>(D)) {
3589     Body = CD->getBody();
3590     NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP;
3591     if (NeedToDelayGlobalization &&
3592         getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
3593       return;
3594   }
3595   if (!Body)
3596     return;
3597   CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second);
3598   VarChecker.Visit(Body);
3599   const RecordDecl *GlobalizedVarsRecord =
3600       VarChecker.getGlobalizedRecord(IsInTTDRegion);
3601   TeamAndReductions.first = nullptr;
3602   TeamAndReductions.second.clear();
3603   ArrayRef<const ValueDecl *> EscapedVariableLengthDecls =
3604       VarChecker.getEscapedVariableLengthDecls();
3605   if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty())
3606     return;
3607   auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
3608   I->getSecond().MappedParams =
3609       std::make_unique<CodeGenFunction::OMPMapVars>();
3610   I->getSecond().EscapedParameters.insert(
3611       VarChecker.getEscapedParameters().begin(),
3612       VarChecker.getEscapedParameters().end());
3613   I->getSecond().EscapedVariableLengthDecls.append(
3614       EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end());
3615   DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
3616   for (const ValueDecl *VD : VarChecker.getEscapedDecls()) {
3617     assert(VD->isCanonicalDecl() && "Expected canonical declaration");
3618     Data.insert(std::make_pair(VD, MappedVarData()));
3619   }
3620   if (!IsInTTDRegion && !NeedToDelayGlobalization && !IsInParallelRegion) {
3621     CheckVarsEscapingDeclContext VarChecker(CGF, llvm::None);
3622     VarChecker.Visit(Body);
3623     I->getSecond().SecondaryLocalVarData.emplace();
3624     DeclToAddrMapTy &Data = I->getSecond().SecondaryLocalVarData.getValue();
3625     for (const ValueDecl *VD : VarChecker.getEscapedDecls()) {
3626       assert(VD->isCanonicalDecl() && "Expected canonical declaration");
3627       Data.insert(std::make_pair(VD, MappedVarData()));
3628     }
3629   }
3630   if (!NeedToDelayGlobalization) {
3631     emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true);
3632     struct GlobalizationScope final : EHScopeStack::Cleanup {
3633       GlobalizationScope() = default;
3634 
3635       void Emit(CodeGenFunction &CGF, Flags flags) override {
3636         static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
3637             .emitGenericVarsEpilog(CGF, /*WithSPMDCheck=*/true);
3638       }
3639     };
3640     CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup);
3641   }
3642 }
3643 
3644 Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF,
3645                                                         const VarDecl *VD) {
3646   if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) {
3647     const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
3648     auto AS = LangAS::Default;
3649     switch (A->getAllocatorType()) {
3650       // Use the default allocator here as by default local vars are
3651       // threadlocal.
3652     case OMPAllocateDeclAttr::OMPNullMemAlloc:
3653     case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
3654     case OMPAllocateDeclAttr::OMPThreadMemAlloc:
3655     case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
3656     case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
3657       // Follow the user decision - use default allocation.
3658       return Address::invalid();
3659     case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
3660       // TODO: implement aupport for user-defined allocators.
3661       return Address::invalid();
3662     case OMPAllocateDeclAttr::OMPConstMemAlloc:
3663       AS = LangAS::cuda_constant;
3664       break;
3665     case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
3666       AS = LangAS::cuda_shared;
3667       break;
3668     case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
3669     case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
3670       break;
3671     }
3672     llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
3673     auto *GV = new llvm::GlobalVariable(
3674         CGM.getModule(), VarTy, /*isConstant=*/false,
3675         llvm::GlobalValue::InternalLinkage, llvm::Constant::getNullValue(VarTy),
3676         VD->getName(),
3677         /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
3678         CGM.getContext().getTargetAddressSpace(AS));
3679     CharUnits Align = CGM.getContext().getDeclAlign(VD);
3680     GV->setAlignment(Align.getAsAlign());
3681     return Address(
3682         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3683             GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace(
3684                     VD->getType().getAddressSpace()))),
3685         Align);
3686   }
3687 
3688   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic)
3689     return Address::invalid();
3690 
3691   VD = VD->getCanonicalDecl();
3692   auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
3693   if (I == FunctionGlobalizedDecls.end())
3694     return Address::invalid();
3695   auto VDI = I->getSecond().LocalVarData.find(VD);
3696   if (VDI != I->getSecond().LocalVarData.end())
3697     return VDI->second.PrivateAddr;
3698   if (VD->hasAttrs()) {
3699     for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()),
3700          E(VD->attr_end());
3701          IT != E; ++IT) {
3702       auto VDI = I->getSecond().LocalVarData.find(
3703           cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl())
3704               ->getCanonicalDecl());
3705       if (VDI != I->getSecond().LocalVarData.end())
3706         return VDI->second.PrivateAddr;
3707     }
3708   }
3709 
3710   return Address::invalid();
3711 }
3712 
3713 void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) {
3714   FunctionGlobalizedDecls.erase(CGF.CurFn);
3715   CGOpenMPRuntime::functionFinished(CGF);
3716 }
3717 
3718 void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk(
3719     CodeGenFunction &CGF, const OMPLoopDirective &S,
3720     OpenMPDistScheduleClauseKind &ScheduleKind,
3721     llvm::Value *&Chunk) const {
3722   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
3723   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
3724     ScheduleKind = OMPC_DIST_SCHEDULE_static;
3725     Chunk = CGF.EmitScalarConversion(
3726         RT.getGPUNumThreads(CGF),
3727         CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
3728         S.getIterationVariable()->getType(), S.getBeginLoc());
3729     return;
3730   }
3731   CGOpenMPRuntime::getDefaultDistScheduleAndChunk(
3732       CGF, S, ScheduleKind, Chunk);
3733 }
3734 
3735 void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk(
3736     CodeGenFunction &CGF, const OMPLoopDirective &S,
3737     OpenMPScheduleClauseKind &ScheduleKind,
3738     const Expr *&ChunkExpr) const {
3739   ScheduleKind = OMPC_SCHEDULE_static;
3740   // Chunk size is 1 in this case.
3741   llvm::APInt ChunkSize(32, 1);
3742   ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize,
3743       CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
3744       SourceLocation());
3745 }
3746 
3747 void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas(
3748     CodeGenFunction &CGF, const OMPExecutableDirective &D) const {
3749   assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&
3750          " Expected target-based directive.");
3751   const CapturedStmt *CS = D.getCapturedStmt(OMPD_target);
3752   for (const CapturedStmt::Capture &C : CS->captures()) {
3753     // Capture variables captured by reference in lambdas for target-based
3754     // directives.
3755     if (!C.capturesVariable())
3756       continue;
3757     const VarDecl *VD = C.getCapturedVar();
3758     const auto *RD = VD->getType()
3759                          .getCanonicalType()
3760                          .getNonReferenceType()
3761                          ->getAsCXXRecordDecl();
3762     if (!RD || !RD->isLambda())
3763       continue;
3764     Address VDAddr = CGF.GetAddrOfLocalVar(VD);
3765     LValue VDLVal;
3766     if (VD->getType().getCanonicalType()->isReferenceType())
3767       VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType());
3768     else
3769       VDLVal = CGF.MakeAddrLValue(
3770           VDAddr, VD->getType().getCanonicalType().getNonReferenceType());
3771     llvm::DenseMap<const VarDecl *, FieldDecl *> Captures;
3772     FieldDecl *ThisCapture = nullptr;
3773     RD->getCaptureFields(Captures, ThisCapture);
3774     if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) {
3775       LValue ThisLVal =
3776           CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture);
3777       llvm::Value *CXXThis = CGF.LoadCXXThis();
3778       CGF.EmitStoreOfScalar(CXXThis, ThisLVal);
3779     }
3780     for (const LambdaCapture &LC : RD->captures()) {
3781       if (LC.getCaptureKind() != LCK_ByRef)
3782         continue;
3783       const VarDecl *VD = LC.getCapturedVar();
3784       if (!CS->capturesVariable(VD))
3785         continue;
3786       auto It = Captures.find(VD);
3787       assert(It != Captures.end() && "Found lambda capture without field.");
3788       LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second);
3789       Address VDAddr = CGF.GetAddrOfLocalVar(VD);
3790       if (VD->getType().getCanonicalType()->isReferenceType())
3791         VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr,
3792                                                VD->getType().getCanonicalType())
3793                      .getAddress(CGF);
3794       CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal);
3795     }
3796   }
3797 }
3798 
3799 bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD,
3800                                                             LangAS &AS) {
3801   if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())
3802     return false;
3803   const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
3804   switch(A->getAllocatorType()) {
3805   case OMPAllocateDeclAttr::OMPNullMemAlloc:
3806   case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
3807   // Not supported, fallback to the default mem space.
3808   case OMPAllocateDeclAttr::OMPThreadMemAlloc:
3809   case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
3810   case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
3811   case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
3812   case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
3813     AS = LangAS::Default;
3814     return true;
3815   case OMPAllocateDeclAttr::OMPConstMemAlloc:
3816     AS = LangAS::cuda_constant;
3817     return true;
3818   case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
3819     AS = LangAS::cuda_shared;
3820     return true;
3821   case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
3822     llvm_unreachable("Expected predefined allocator for the variables with the "
3823                      "static storage.");
3824   }
3825   return false;
3826 }
3827 
3828 // Get current CudaArch and ignore any unknown values
3829 static CudaArch getCudaArch(CodeGenModule &CGM) {
3830   if (!CGM.getTarget().hasFeature("ptx"))
3831     return CudaArch::UNKNOWN;
3832   for (const auto &Feature : CGM.getTarget().getTargetOpts().FeatureMap) {
3833     if (Feature.getValue()) {
3834       CudaArch Arch = StringToCudaArch(Feature.getKey());
3835       if (Arch != CudaArch::UNKNOWN)
3836         return Arch;
3837     }
3838   }
3839   return CudaArch::UNKNOWN;
3840 }
3841 
3842 /// Check to see if target architecture supports unified addressing which is
3843 /// a restriction for OpenMP requires clause "unified_shared_memory".
3844 void CGOpenMPRuntimeGPU::processRequiresDirective(
3845     const OMPRequiresDecl *D) {
3846   for (const OMPClause *Clause : D->clauselists()) {
3847     if (Clause->getClauseKind() == OMPC_unified_shared_memory) {
3848       CudaArch Arch = getCudaArch(CGM);
3849       switch (Arch) {
3850       case CudaArch::SM_20:
3851       case CudaArch::SM_21:
3852       case CudaArch::SM_30:
3853       case CudaArch::SM_32:
3854       case CudaArch::SM_35:
3855       case CudaArch::SM_37:
3856       case CudaArch::SM_50:
3857       case CudaArch::SM_52:
3858       case CudaArch::SM_53: {
3859         SmallString<256> Buffer;
3860         llvm::raw_svector_ostream Out(Buffer);
3861         Out << "Target architecture " << CudaArchToString(Arch)
3862             << " does not support unified addressing";
3863         CGM.Error(Clause->getBeginLoc(), Out.str());
3864         return;
3865       }
3866       case CudaArch::SM_60:
3867       case CudaArch::SM_61:
3868       case CudaArch::SM_62:
3869       case CudaArch::SM_70:
3870       case CudaArch::SM_72:
3871       case CudaArch::SM_75:
3872       case CudaArch::SM_80:
3873       case CudaArch::SM_86:
3874       case CudaArch::GFX600:
3875       case CudaArch::GFX601:
3876       case CudaArch::GFX602:
3877       case CudaArch::GFX700:
3878       case CudaArch::GFX701:
3879       case CudaArch::GFX702:
3880       case CudaArch::GFX703:
3881       case CudaArch::GFX704:
3882       case CudaArch::GFX705:
3883       case CudaArch::GFX801:
3884       case CudaArch::GFX802:
3885       case CudaArch::GFX803:
3886       case CudaArch::GFX805:
3887       case CudaArch::GFX810:
3888       case CudaArch::GFX900:
3889       case CudaArch::GFX902:
3890       case CudaArch::GFX904:
3891       case CudaArch::GFX906:
3892       case CudaArch::GFX908:
3893       case CudaArch::GFX909:
3894       case CudaArch::GFX90a:
3895       case CudaArch::GFX90c:
3896       case CudaArch::GFX1010:
3897       case CudaArch::GFX1011:
3898       case CudaArch::GFX1012:
3899       case CudaArch::GFX1013:
3900       case CudaArch::GFX1030:
3901       case CudaArch::GFX1031:
3902       case CudaArch::GFX1032:
3903       case CudaArch::GFX1033:
3904       case CudaArch::GFX1034:
3905       case CudaArch::GFX1035:
3906       case CudaArch::Generic:
3907       case CudaArch::UNUSED:
3908       case CudaArch::UNKNOWN:
3909         break;
3910       case CudaArch::LAST:
3911         llvm_unreachable("Unexpected Cuda arch.");
3912       }
3913     }
3914   }
3915   CGOpenMPRuntime::processRequiresDirective(D);
3916 }
3917 
3918 void CGOpenMPRuntimeGPU::clear() {
3919 
3920   if (!TeamsReductions.empty()) {
3921     ASTContext &C = CGM.getContext();
3922     RecordDecl *StaticRD = C.buildImplicitRecord(
3923         "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::TTK_Union);
3924     StaticRD->startDefinition();
3925     for (const RecordDecl *TeamReductionRec : TeamsReductions) {
3926       QualType RecTy = C.getRecordType(TeamReductionRec);
3927       auto *Field = FieldDecl::Create(
3928           C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy,
3929           C.getTrivialTypeSourceInfo(RecTy, SourceLocation()),
3930           /*BW=*/nullptr, /*Mutable=*/false,
3931           /*InitStyle=*/ICIS_NoInit);
3932       Field->setAccess(AS_public);
3933       StaticRD->addDecl(Field);
3934     }
3935     StaticRD->completeDefinition();
3936     QualType StaticTy = C.getRecordType(StaticRD);
3937     llvm::Type *LLVMReductionsBufferTy =
3938         CGM.getTypes().ConvertTypeForMem(StaticTy);
3939     // FIXME: nvlink does not handle weak linkage correctly (object with the
3940     // different size are reported as erroneous).
3941     // Restore CommonLinkage as soon as nvlink is fixed.
3942     auto *GV = new llvm::GlobalVariable(
3943         CGM.getModule(), LLVMReductionsBufferTy,
3944         /*isConstant=*/false, llvm::GlobalValue::InternalLinkage,
3945         llvm::Constant::getNullValue(LLVMReductionsBufferTy),
3946         "_openmp_teams_reductions_buffer_$_");
3947     KernelTeamsReductionPtr->setInitializer(
3948         llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV,
3949                                                              CGM.VoidPtrTy));
3950   }
3951   CGOpenMPRuntime::clear();
3952 }
3953 
3954 llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) {
3955   CGBuilderTy &Bld = CGF.Builder;
3956   llvm::Module *M = &CGF.CGM.getModule();
3957   const char *LocSize = "__kmpc_get_hardware_num_threads_in_block";
3958   llvm::Function *F = M->getFunction(LocSize);
3959   if (!F) {
3960     F = llvm::Function::Create(
3961         llvm::FunctionType::get(CGF.Int32Ty, llvm::None, false),
3962         llvm::GlobalVariable::ExternalLinkage, LocSize, &CGF.CGM.getModule());
3963   }
3964   return Bld.CreateCall(F, llvm::None, "nvptx_num_threads");
3965 }
3966 
3967 llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) {
3968   ArrayRef<llvm::Value *> Args{};
3969   return CGF.EmitRuntimeCall(
3970       OMPBuilder.getOrCreateRuntimeFunction(
3971           CGM.getModule(), OMPRTL___kmpc_get_hardware_thread_id_in_block),
3972       Args);
3973 }
3974 
3975 llvm::Value *CGOpenMPRuntimeGPU::getGPUWarpSize(CodeGenFunction &CGF) {
3976   ArrayRef<llvm::Value *> Args{};
3977   return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
3978                                  CGM.getModule(), OMPRTL___kmpc_get_warp_size),
3979                              Args);
3980 }
3981