1 //===---- CGOpenMPRuntimeGPU.cpp - Interface to OpenMP GPU Runtimes ----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This provides a generalized class for OpenMP runtime code generation
10 // specialized by GPU targets NVPTX and AMDGCN.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "CGOpenMPRuntimeGPU.h"
15 #include "CodeGenFunction.h"
16 #include "clang/AST/Attr.h"
17 #include "clang/AST/DeclOpenMP.h"
18 #include "clang/AST/OpenMPClause.h"
19 #include "clang/AST/StmtOpenMP.h"
20 #include "clang/AST/StmtVisitor.h"
21 #include "clang/Basic/Cuda.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/Frontend/OpenMP/OMPGridValues.h"
24 #include "llvm/Support/MathExtras.h"
25 
26 using namespace clang;
27 using namespace CodeGen;
28 using namespace llvm::omp;
29 
30 namespace {
31 /// Pre(post)-action for different OpenMP constructs specialized for NVPTX.
32 class NVPTXActionTy final : public PrePostActionTy {
33   llvm::FunctionCallee EnterCallee = nullptr;
34   ArrayRef<llvm::Value *> EnterArgs;
35   llvm::FunctionCallee ExitCallee = nullptr;
36   ArrayRef<llvm::Value *> ExitArgs;
37   bool Conditional = false;
38   llvm::BasicBlock *ContBlock = nullptr;
39 
40 public:
41   NVPTXActionTy(llvm::FunctionCallee EnterCallee,
42                 ArrayRef<llvm::Value *> EnterArgs,
43                 llvm::FunctionCallee ExitCallee,
44                 ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false)
45       : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee),
46         ExitArgs(ExitArgs), Conditional(Conditional) {}
47   void Enter(CodeGenFunction &CGF) override {
48     llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs);
49     if (Conditional) {
50       llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes);
51       auto *ThenBlock = CGF.createBasicBlock("omp_if.then");
52       ContBlock = CGF.createBasicBlock("omp_if.end");
53       // Generate the branch (If-stmt)
54       CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock);
55       CGF.EmitBlock(ThenBlock);
56     }
57   }
58   void Done(CodeGenFunction &CGF) {
59     // Emit the rest of blocks/branches
60     CGF.EmitBranch(ContBlock);
61     CGF.EmitBlock(ContBlock, true);
62   }
63   void Exit(CodeGenFunction &CGF) override {
64     CGF.EmitRuntimeCall(ExitCallee, ExitArgs);
65   }
66 };
67 
68 /// A class to track the execution mode when codegening directives within
69 /// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry
70 /// to the target region and used by containing directives such as 'parallel'
71 /// to emit optimized code.
72 class ExecutionRuntimeModesRAII {
73 private:
74   CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode =
75       CGOpenMPRuntimeGPU::EM_Unknown;
76   CGOpenMPRuntimeGPU::ExecutionMode &ExecMode;
77 
78 public:
79   ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode,
80                             CGOpenMPRuntimeGPU::ExecutionMode EntryMode)
81       : ExecMode(ExecMode) {
82     SavedExecMode = ExecMode;
83     ExecMode = EntryMode;
84   }
85   ~ExecutionRuntimeModesRAII() { ExecMode = SavedExecMode; }
86 };
87 
88 /// GPU Configuration:  This information can be derived from cuda registers,
89 /// however, providing compile time constants helps generate more efficient
90 /// code.  For all practical purposes this is fine because the configuration
91 /// is the same for all known NVPTX architectures.
92 enum MachineConfiguration : unsigned {
93   /// See "llvm/Frontend/OpenMP/OMPGridValues.h" for various related target
94   /// specific Grid Values like GV_Warp_Size, GV_Slot_Size
95 
96   /// Global memory alignment for performance.
97   GlobalMemoryAlignment = 128,
98 };
99 
100 static const ValueDecl *getPrivateItem(const Expr *RefExpr) {
101   RefExpr = RefExpr->IgnoreParens();
102   if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) {
103     const Expr *Base = ASE->getBase()->IgnoreParenImpCasts();
104     while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
105       Base = TempASE->getBase()->IgnoreParenImpCasts();
106     RefExpr = Base;
107   } else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) {
108     const Expr *Base = OASE->getBase()->IgnoreParenImpCasts();
109     while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base))
110       Base = TempOASE->getBase()->IgnoreParenImpCasts();
111     while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base))
112       Base = TempASE->getBase()->IgnoreParenImpCasts();
113     RefExpr = Base;
114   }
115   RefExpr = RefExpr->IgnoreParenImpCasts();
116   if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr))
117     return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl());
118   const auto *ME = cast<MemberExpr>(RefExpr);
119   return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl());
120 }
121 
122 
123 static RecordDecl *buildRecordForGlobalizedVars(
124     ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls,
125     ArrayRef<const ValueDecl *> EscapedDeclsForTeams,
126     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
127         &MappedDeclsFields, int BufSize) {
128   using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>;
129   if (EscapedDecls.empty() && EscapedDeclsForTeams.empty())
130     return nullptr;
131   SmallVector<VarsDataTy, 4> GlobalizedVars;
132   for (const ValueDecl *D : EscapedDecls)
133     GlobalizedVars.emplace_back(
134         CharUnits::fromQuantity(std::max(
135             C.getDeclAlign(D).getQuantity(),
136             static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))),
137         D);
138   for (const ValueDecl *D : EscapedDeclsForTeams)
139     GlobalizedVars.emplace_back(C.getDeclAlign(D), D);
140   llvm::stable_sort(GlobalizedVars, [](VarsDataTy L, VarsDataTy R) {
141     return L.first > R.first;
142   });
143 
144   // Build struct _globalized_locals_ty {
145   //         /*  globalized vars  */[WarSize] align (max(decl_align,
146   //         GlobalMemoryAlignment))
147   //         /*  globalized vars  */ for EscapedDeclsForTeams
148   //       };
149   RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty");
150   GlobalizedRD->startDefinition();
151   llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped(
152       EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end());
153   for (const auto &Pair : GlobalizedVars) {
154     const ValueDecl *VD = Pair.second;
155     QualType Type = VD->getType();
156     if (Type->isLValueReferenceType())
157       Type = C.getPointerType(Type.getNonReferenceType());
158     else
159       Type = Type.getNonReferenceType();
160     SourceLocation Loc = VD->getLocation();
161     FieldDecl *Field;
162     if (SingleEscaped.count(VD)) {
163       Field = FieldDecl::Create(
164           C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
165           C.getTrivialTypeSourceInfo(Type, SourceLocation()),
166           /*BW=*/nullptr, /*Mutable=*/false,
167           /*InitStyle=*/ICIS_NoInit);
168       Field->setAccess(AS_public);
169       if (VD->hasAttrs()) {
170         for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
171              E(VD->getAttrs().end());
172              I != E; ++I)
173           Field->addAttr(*I);
174       }
175     } else {
176       llvm::APInt ArraySize(32, BufSize);
177       Type = C.getConstantArrayType(Type, ArraySize, nullptr, ArrayType::Normal,
178                                     0);
179       Field = FieldDecl::Create(
180           C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
181           C.getTrivialTypeSourceInfo(Type, SourceLocation()),
182           /*BW=*/nullptr, /*Mutable=*/false,
183           /*InitStyle=*/ICIS_NoInit);
184       Field->setAccess(AS_public);
185       llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(),
186                                      static_cast<CharUnits::QuantityType>(
187                                          GlobalMemoryAlignment)));
188       Field->addAttr(AlignedAttr::CreateImplicit(
189           C, /*IsAlignmentExpr=*/true,
190           IntegerLiteral::Create(C, Align,
191                                  C.getIntTypeForBitwidth(32, /*Signed=*/0),
192                                  SourceLocation()),
193           {}, AttributeCommonInfo::AS_GNU, AlignedAttr::GNU_aligned));
194     }
195     GlobalizedRD->addDecl(Field);
196     MappedDeclsFields.try_emplace(VD, Field);
197   }
198   GlobalizedRD->completeDefinition();
199   return GlobalizedRD;
200 }
201 
202 /// Get the list of variables that can escape their declaration context.
203 class CheckVarsEscapingDeclContext final
204     : public ConstStmtVisitor<CheckVarsEscapingDeclContext> {
205   CodeGenFunction &CGF;
206   llvm::SetVector<const ValueDecl *> EscapedDecls;
207   llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls;
208   llvm::SmallPtrSet<const Decl *, 4> EscapedParameters;
209   RecordDecl *GlobalizedRD = nullptr;
210   llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
211   bool AllEscaped = false;
212   bool IsForCombinedParallelRegion = false;
213 
214   void markAsEscaped(const ValueDecl *VD) {
215     // Do not globalize declare target variables.
216     if (!isa<VarDecl>(VD) ||
217         OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD))
218       return;
219     VD = cast<ValueDecl>(VD->getCanonicalDecl());
220     // Use user-specified allocation.
221     if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>())
222       return;
223     // Variables captured by value must be globalized.
224     if (auto *CSI = CGF.CapturedStmtInfo) {
225       if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) {
226         // Check if need to capture the variable that was already captured by
227         // value in the outer region.
228         if (!IsForCombinedParallelRegion) {
229           if (!FD->hasAttrs())
230             return;
231           const auto *Attr = FD->getAttr<OMPCaptureKindAttr>();
232           if (!Attr)
233             return;
234           if (((Attr->getCaptureKind() != OMPC_map) &&
235                !isOpenMPPrivate(Attr->getCaptureKind())) ||
236               ((Attr->getCaptureKind() == OMPC_map) &&
237                !FD->getType()->isAnyPointerType()))
238             return;
239         }
240         if (!FD->getType()->isReferenceType()) {
241           assert(!VD->getType()->isVariablyModifiedType() &&
242                  "Parameter captured by value with variably modified type");
243           EscapedParameters.insert(VD);
244         } else if (!IsForCombinedParallelRegion) {
245           return;
246         }
247       }
248     }
249     if ((!CGF.CapturedStmtInfo ||
250          (IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) &&
251         VD->getType()->isReferenceType())
252       // Do not globalize variables with reference type.
253       return;
254     if (VD->getType()->isVariablyModifiedType())
255       EscapedVariableLengthDecls.insert(VD);
256     else
257       EscapedDecls.insert(VD);
258   }
259 
260   void VisitValueDecl(const ValueDecl *VD) {
261     if (VD->getType()->isLValueReferenceType())
262       markAsEscaped(VD);
263     if (const auto *VarD = dyn_cast<VarDecl>(VD)) {
264       if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) {
265         const bool SavedAllEscaped = AllEscaped;
266         AllEscaped = VD->getType()->isLValueReferenceType();
267         Visit(VarD->getInit());
268         AllEscaped = SavedAllEscaped;
269       }
270     }
271   }
272   void VisitOpenMPCapturedStmt(const CapturedStmt *S,
273                                ArrayRef<OMPClause *> Clauses,
274                                bool IsCombinedParallelRegion) {
275     if (!S)
276       return;
277     for (const CapturedStmt::Capture &C : S->captures()) {
278       if (C.capturesVariable() && !C.capturesVariableByCopy()) {
279         const ValueDecl *VD = C.getCapturedVar();
280         bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion;
281         if (IsCombinedParallelRegion) {
282           // Check if the variable is privatized in the combined construct and
283           // those private copies must be shared in the inner parallel
284           // directive.
285           IsForCombinedParallelRegion = false;
286           for (const OMPClause *C : Clauses) {
287             if (!isOpenMPPrivate(C->getClauseKind()) ||
288                 C->getClauseKind() == OMPC_reduction ||
289                 C->getClauseKind() == OMPC_linear ||
290                 C->getClauseKind() == OMPC_private)
291               continue;
292             ArrayRef<const Expr *> Vars;
293             if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C))
294               Vars = PC->getVarRefs();
295             else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C))
296               Vars = PC->getVarRefs();
297             else
298               llvm_unreachable("Unexpected clause.");
299             for (const auto *E : Vars) {
300               const Decl *D =
301                   cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl();
302               if (D == VD->getCanonicalDecl()) {
303                 IsForCombinedParallelRegion = true;
304                 break;
305               }
306             }
307             if (IsForCombinedParallelRegion)
308               break;
309           }
310         }
311         markAsEscaped(VD);
312         if (isa<OMPCapturedExprDecl>(VD))
313           VisitValueDecl(VD);
314         IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion;
315       }
316     }
317   }
318 
319   void buildRecordForGlobalizedVars(bool IsInTTDRegion) {
320     assert(!GlobalizedRD &&
321            "Record for globalized variables is built already.");
322     ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams;
323     unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size;
324     if (IsInTTDRegion)
325       EscapedDeclsForTeams = EscapedDecls.getArrayRef();
326     else
327       EscapedDeclsForParallel = EscapedDecls.getArrayRef();
328     GlobalizedRD = ::buildRecordForGlobalizedVars(
329         CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams,
330         MappedDeclsFields, WarpSize);
331   }
332 
333 public:
334   CheckVarsEscapingDeclContext(CodeGenFunction &CGF,
335                                ArrayRef<const ValueDecl *> TeamsReductions)
336       : CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) {
337   }
338   virtual ~CheckVarsEscapingDeclContext() = default;
339   void VisitDeclStmt(const DeclStmt *S) {
340     if (!S)
341       return;
342     for (const Decl *D : S->decls())
343       if (const auto *VD = dyn_cast_or_null<ValueDecl>(D))
344         VisitValueDecl(VD);
345   }
346   void VisitOMPExecutableDirective(const OMPExecutableDirective *D) {
347     if (!D)
348       return;
349     if (!D->hasAssociatedStmt())
350       return;
351     if (const auto *S =
352             dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) {
353       // Do not analyze directives that do not actually require capturing,
354       // like `omp for` or `omp simd` directives.
355       llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
356       getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind());
357       if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) {
358         VisitStmt(S->getCapturedStmt());
359         return;
360       }
361       VisitOpenMPCapturedStmt(
362           S, D->clauses(),
363           CaptureRegions.back() == OMPD_parallel &&
364               isOpenMPDistributeDirective(D->getDirectiveKind()));
365     }
366   }
367   void VisitCapturedStmt(const CapturedStmt *S) {
368     if (!S)
369       return;
370     for (const CapturedStmt::Capture &C : S->captures()) {
371       if (C.capturesVariable() && !C.capturesVariableByCopy()) {
372         const ValueDecl *VD = C.getCapturedVar();
373         markAsEscaped(VD);
374         if (isa<OMPCapturedExprDecl>(VD))
375           VisitValueDecl(VD);
376       }
377     }
378   }
379   void VisitLambdaExpr(const LambdaExpr *E) {
380     if (!E)
381       return;
382     for (const LambdaCapture &C : E->captures()) {
383       if (C.capturesVariable()) {
384         if (C.getCaptureKind() == LCK_ByRef) {
385           const ValueDecl *VD = C.getCapturedVar();
386           markAsEscaped(VD);
387           if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD))
388             VisitValueDecl(VD);
389         }
390       }
391     }
392   }
393   void VisitBlockExpr(const BlockExpr *E) {
394     if (!E)
395       return;
396     for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) {
397       if (C.isByRef()) {
398         const VarDecl *VD = C.getVariable();
399         markAsEscaped(VD);
400         if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture())
401           VisitValueDecl(VD);
402       }
403     }
404   }
405   void VisitCallExpr(const CallExpr *E) {
406     if (!E)
407       return;
408     for (const Expr *Arg : E->arguments()) {
409       if (!Arg)
410         continue;
411       if (Arg->isLValue()) {
412         const bool SavedAllEscaped = AllEscaped;
413         AllEscaped = true;
414         Visit(Arg);
415         AllEscaped = SavedAllEscaped;
416       } else {
417         Visit(Arg);
418       }
419     }
420     Visit(E->getCallee());
421   }
422   void VisitDeclRefExpr(const DeclRefExpr *E) {
423     if (!E)
424       return;
425     const ValueDecl *VD = E->getDecl();
426     if (AllEscaped)
427       markAsEscaped(VD);
428     if (isa<OMPCapturedExprDecl>(VD))
429       VisitValueDecl(VD);
430     else if (VD->isInitCapture())
431       VisitValueDecl(VD);
432   }
433   void VisitUnaryOperator(const UnaryOperator *E) {
434     if (!E)
435       return;
436     if (E->getOpcode() == UO_AddrOf) {
437       const bool SavedAllEscaped = AllEscaped;
438       AllEscaped = true;
439       Visit(E->getSubExpr());
440       AllEscaped = SavedAllEscaped;
441     } else {
442       Visit(E->getSubExpr());
443     }
444   }
445   void VisitImplicitCastExpr(const ImplicitCastExpr *E) {
446     if (!E)
447       return;
448     if (E->getCastKind() == CK_ArrayToPointerDecay) {
449       const bool SavedAllEscaped = AllEscaped;
450       AllEscaped = true;
451       Visit(E->getSubExpr());
452       AllEscaped = SavedAllEscaped;
453     } else {
454       Visit(E->getSubExpr());
455     }
456   }
457   void VisitExpr(const Expr *E) {
458     if (!E)
459       return;
460     bool SavedAllEscaped = AllEscaped;
461     if (!E->isLValue())
462       AllEscaped = false;
463     for (const Stmt *Child : E->children())
464       if (Child)
465         Visit(Child);
466     AllEscaped = SavedAllEscaped;
467   }
468   void VisitStmt(const Stmt *S) {
469     if (!S)
470       return;
471     for (const Stmt *Child : S->children())
472       if (Child)
473         Visit(Child);
474   }
475 
476   /// Returns the record that handles all the escaped local variables and used
477   /// instead of their original storage.
478   const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) {
479     if (!GlobalizedRD)
480       buildRecordForGlobalizedVars(IsInTTDRegion);
481     return GlobalizedRD;
482   }
483 
484   /// Returns the field in the globalized record for the escaped variable.
485   const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const {
486     assert(GlobalizedRD &&
487            "Record for globalized variables must be generated already.");
488     auto I = MappedDeclsFields.find(VD);
489     if (I == MappedDeclsFields.end())
490       return nullptr;
491     return I->getSecond();
492   }
493 
494   /// Returns the list of the escaped local variables/parameters.
495   ArrayRef<const ValueDecl *> getEscapedDecls() const {
496     return EscapedDecls.getArrayRef();
497   }
498 
499   /// Checks if the escaped local variable is actually a parameter passed by
500   /// value.
501   const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const {
502     return EscapedParameters;
503   }
504 
505   /// Returns the list of the escaped variables with the variably modified
506   /// types.
507   ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const {
508     return EscapedVariableLengthDecls.getArrayRef();
509   }
510 };
511 } // anonymous namespace
512 
513 /// Get the id of the warp in the block.
514 /// We assume that the warp size is 32, which is always the case
515 /// on the NVPTX device, to generate more efficient code.
516 static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) {
517   CGBuilderTy &Bld = CGF.Builder;
518   unsigned LaneIDBits =
519       llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size);
520   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
521   return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id");
522 }
523 
524 /// Get the id of the current lane in the Warp.
525 /// We assume that the warp size is 32, which is always the case
526 /// on the NVPTX device, to generate more efficient code.
527 static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) {
528   CGBuilderTy &Bld = CGF.Builder;
529   unsigned LaneIDBits =
530       llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size);
531   unsigned LaneIDMask = ~0u >> (32u - LaneIDBits);
532   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
533   return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask),
534                        "nvptx_lane_id");
535 }
536 
537 CGOpenMPRuntimeGPU::ExecutionMode
538 CGOpenMPRuntimeGPU::getExecutionMode() const {
539   return CurrentExecutionMode;
540 }
541 
542 static CGOpenMPRuntimeGPU::DataSharingMode
543 getDataSharingMode(CodeGenModule &CGM) {
544   return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeGPU::CUDA
545                                           : CGOpenMPRuntimeGPU::Generic;
546 }
547 
548 /// Check for inner (nested) SPMD construct, if any
549 static bool hasNestedSPMDDirective(ASTContext &Ctx,
550                                    const OMPExecutableDirective &D) {
551   const auto *CS = D.getInnermostCapturedStmt();
552   const auto *Body =
553       CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true);
554   const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
555 
556   if (const auto *NestedDir =
557           dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
558     OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
559     switch (D.getDirectiveKind()) {
560     case OMPD_target:
561       if (isOpenMPParallelDirective(DKind))
562         return true;
563       if (DKind == OMPD_teams) {
564         Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
565             /*IgnoreCaptured=*/true);
566         if (!Body)
567           return false;
568         ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body);
569         if (const auto *NND =
570                 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
571           DKind = NND->getDirectiveKind();
572           if (isOpenMPParallelDirective(DKind))
573             return true;
574         }
575       }
576       return false;
577     case OMPD_target_teams:
578       return isOpenMPParallelDirective(DKind);
579     case OMPD_target_simd:
580     case OMPD_target_parallel:
581     case OMPD_target_parallel_for:
582     case OMPD_target_parallel_for_simd:
583     case OMPD_target_teams_distribute:
584     case OMPD_target_teams_distribute_simd:
585     case OMPD_target_teams_distribute_parallel_for:
586     case OMPD_target_teams_distribute_parallel_for_simd:
587     case OMPD_parallel:
588     case OMPD_for:
589     case OMPD_parallel_for:
590     case OMPD_parallel_master:
591     case OMPD_parallel_sections:
592     case OMPD_for_simd:
593     case OMPD_parallel_for_simd:
594     case OMPD_cancel:
595     case OMPD_cancellation_point:
596     case OMPD_ordered:
597     case OMPD_threadprivate:
598     case OMPD_allocate:
599     case OMPD_task:
600     case OMPD_simd:
601     case OMPD_sections:
602     case OMPD_section:
603     case OMPD_single:
604     case OMPD_master:
605     case OMPD_critical:
606     case OMPD_taskyield:
607     case OMPD_barrier:
608     case OMPD_taskwait:
609     case OMPD_taskgroup:
610     case OMPD_atomic:
611     case OMPD_flush:
612     case OMPD_depobj:
613     case OMPD_scan:
614     case OMPD_teams:
615     case OMPD_target_data:
616     case OMPD_target_exit_data:
617     case OMPD_target_enter_data:
618     case OMPD_distribute:
619     case OMPD_distribute_simd:
620     case OMPD_distribute_parallel_for:
621     case OMPD_distribute_parallel_for_simd:
622     case OMPD_teams_distribute:
623     case OMPD_teams_distribute_simd:
624     case OMPD_teams_distribute_parallel_for:
625     case OMPD_teams_distribute_parallel_for_simd:
626     case OMPD_target_update:
627     case OMPD_declare_simd:
628     case OMPD_declare_variant:
629     case OMPD_begin_declare_variant:
630     case OMPD_end_declare_variant:
631     case OMPD_declare_target:
632     case OMPD_end_declare_target:
633     case OMPD_declare_reduction:
634     case OMPD_declare_mapper:
635     case OMPD_taskloop:
636     case OMPD_taskloop_simd:
637     case OMPD_master_taskloop:
638     case OMPD_master_taskloop_simd:
639     case OMPD_parallel_master_taskloop:
640     case OMPD_parallel_master_taskloop_simd:
641     case OMPD_requires:
642     case OMPD_unknown:
643     default:
644       llvm_unreachable("Unexpected directive.");
645     }
646   }
647 
648   return false;
649 }
650 
651 static bool supportsSPMDExecutionMode(ASTContext &Ctx,
652                                       const OMPExecutableDirective &D) {
653   OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind();
654   switch (DirectiveKind) {
655   case OMPD_target:
656   case OMPD_target_teams:
657     return hasNestedSPMDDirective(Ctx, D);
658   case OMPD_target_parallel:
659   case OMPD_target_parallel_for:
660   case OMPD_target_parallel_for_simd:
661   case OMPD_target_teams_distribute_parallel_for:
662   case OMPD_target_teams_distribute_parallel_for_simd:
663   case OMPD_target_simd:
664   case OMPD_target_teams_distribute_simd:
665     return true;
666   case OMPD_target_teams_distribute:
667     return false;
668   case OMPD_parallel:
669   case OMPD_for:
670   case OMPD_parallel_for:
671   case OMPD_parallel_master:
672   case OMPD_parallel_sections:
673   case OMPD_for_simd:
674   case OMPD_parallel_for_simd:
675   case OMPD_cancel:
676   case OMPD_cancellation_point:
677   case OMPD_ordered:
678   case OMPD_threadprivate:
679   case OMPD_allocate:
680   case OMPD_task:
681   case OMPD_simd:
682   case OMPD_sections:
683   case OMPD_section:
684   case OMPD_single:
685   case OMPD_master:
686   case OMPD_critical:
687   case OMPD_taskyield:
688   case OMPD_barrier:
689   case OMPD_taskwait:
690   case OMPD_taskgroup:
691   case OMPD_atomic:
692   case OMPD_flush:
693   case OMPD_depobj:
694   case OMPD_scan:
695   case OMPD_teams:
696   case OMPD_target_data:
697   case OMPD_target_exit_data:
698   case OMPD_target_enter_data:
699   case OMPD_distribute:
700   case OMPD_distribute_simd:
701   case OMPD_distribute_parallel_for:
702   case OMPD_distribute_parallel_for_simd:
703   case OMPD_teams_distribute:
704   case OMPD_teams_distribute_simd:
705   case OMPD_teams_distribute_parallel_for:
706   case OMPD_teams_distribute_parallel_for_simd:
707   case OMPD_target_update:
708   case OMPD_declare_simd:
709   case OMPD_declare_variant:
710   case OMPD_begin_declare_variant:
711   case OMPD_end_declare_variant:
712   case OMPD_declare_target:
713   case OMPD_end_declare_target:
714   case OMPD_declare_reduction:
715   case OMPD_declare_mapper:
716   case OMPD_taskloop:
717   case OMPD_taskloop_simd:
718   case OMPD_master_taskloop:
719   case OMPD_master_taskloop_simd:
720   case OMPD_parallel_master_taskloop:
721   case OMPD_parallel_master_taskloop_simd:
722   case OMPD_requires:
723   case OMPD_unknown:
724   default:
725     break;
726   }
727   llvm_unreachable(
728       "Unknown programming model for OpenMP directive on NVPTX target.");
729 }
730 
731 void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D,
732                                              StringRef ParentName,
733                                              llvm::Function *&OutlinedFn,
734                                              llvm::Constant *&OutlinedFnID,
735                                              bool IsOffloadEntry,
736                                              const RegionCodeGenTy &CodeGen) {
737   ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_NonSPMD);
738   EntryFunctionState EST;
739   WrapperFunctionsMap.clear();
740 
741   // Emit target region as a standalone region.
742   class NVPTXPrePostActionTy : public PrePostActionTy {
743     CGOpenMPRuntimeGPU::EntryFunctionState &EST;
744 
745   public:
746     NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST)
747         : EST(EST) {}
748     void Enter(CodeGenFunction &CGF) override {
749       auto &RT =
750           static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
751       RT.emitKernelInit(CGF, EST, /* IsSPMD */ false);
752       // Skip target region initialization.
753       RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
754     }
755     void Exit(CodeGenFunction &CGF) override {
756       auto &RT =
757           static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
758       RT.clearLocThreadIdInsertPt(CGF);
759       RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ false);
760     }
761   } Action(EST);
762   CodeGen.setAction(Action);
763   IsInTTDRegion = true;
764   emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
765                                    IsOffloadEntry, CodeGen);
766   IsInTTDRegion = false;
767 }
768 
769 void CGOpenMPRuntimeGPU::emitKernelInit(CodeGenFunction &CGF,
770                                         EntryFunctionState &EST, bool IsSPMD) {
771   CGBuilderTy &Bld = CGF.Builder;
772   Bld.restoreIP(OMPBuilder.createTargetInit(Bld, IsSPMD));
773   if (!IsSPMD)
774     emitGenericVarsProlog(CGF, EST.Loc);
775 }
776 
777 void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF,
778                                           EntryFunctionState &EST,
779                                           bool IsSPMD) {
780   if (!IsSPMD)
781     emitGenericVarsEpilog(CGF);
782 
783   CGBuilderTy &Bld = CGF.Builder;
784   OMPBuilder.createTargetDeinit(Bld, IsSPMD);
785 }
786 
787 void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D,
788                                           StringRef ParentName,
789                                           llvm::Function *&OutlinedFn,
790                                           llvm::Constant *&OutlinedFnID,
791                                           bool IsOffloadEntry,
792                                           const RegionCodeGenTy &CodeGen) {
793   ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_SPMD);
794   EntryFunctionState EST;
795 
796   // Emit target region as a standalone region.
797   class NVPTXPrePostActionTy : public PrePostActionTy {
798     CGOpenMPRuntimeGPU &RT;
799     CGOpenMPRuntimeGPU::EntryFunctionState &EST;
800 
801   public:
802     NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT,
803                          CGOpenMPRuntimeGPU::EntryFunctionState &EST)
804         : RT(RT), EST(EST) {}
805     void Enter(CodeGenFunction &CGF) override {
806       RT.emitKernelInit(CGF, EST, /* IsSPMD */ true);
807       // Skip target region initialization.
808       RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true);
809     }
810     void Exit(CodeGenFunction &CGF) override {
811       RT.clearLocThreadIdInsertPt(CGF);
812       RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true);
813     }
814   } Action(*this, EST);
815   CodeGen.setAction(Action);
816   IsInTTDRegion = true;
817   emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID,
818                                    IsOffloadEntry, CodeGen);
819   IsInTTDRegion = false;
820 }
821 
822 // Create a unique global variable to indicate the execution mode of this target
823 // region. The execution mode is either 'generic', or 'spmd' depending on the
824 // target directive. This variable is picked up by the offload library to setup
825 // the device appropriately before kernel launch. If the execution mode is
826 // 'generic', the runtime reserves one warp for the master, otherwise, all
827 // warps participate in parallel work.
828 static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name,
829                                      bool Mode) {
830   auto *GVMode = new llvm::GlobalVariable(
831       CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true,
832       llvm::GlobalValue::WeakAnyLinkage,
833       llvm::ConstantInt::get(CGM.Int8Ty, Mode ? OMP_TGT_EXEC_MODE_SPMD
834                                               : OMP_TGT_EXEC_MODE_GENERIC),
835       Twine(Name, "_exec_mode"));
836   GVMode->setVisibility(llvm::GlobalVariable::ProtectedVisibility);
837   CGM.addCompilerUsedGlobal(GVMode);
838 }
839 
840 void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction(
841     const OMPExecutableDirective &D, StringRef ParentName,
842     llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID,
843     bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) {
844   if (!IsOffloadEntry) // Nothing to do.
845     return;
846 
847   assert(!ParentName.empty() && "Invalid target region parent name!");
848 
849   bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D);
850   if (Mode)
851     emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
852                    CodeGen);
853   else
854     emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry,
855                       CodeGen);
856 
857   setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode);
858 }
859 
860 CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM)
861     : CGOpenMPRuntime(CGM) {
862   llvm::OpenMPIRBuilderConfig Config(CGM.getLangOpts().OpenMPIsDevice, true,
863                                      hasRequiresUnifiedSharedMemory(),
864                                      CGM.getLangOpts().OpenMPOffloadMandatory);
865   OMPBuilder.setConfig(Config);
866   OffloadEntriesInfoManager.setConfig(Config);
867 
868   if (!CGM.getLangOpts().OpenMPIsDevice)
869     llvm_unreachable("OpenMP can only handle device code.");
870 
871   llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder();
872   if (CGM.getLangOpts().NoGPULib || CGM.getLangOpts().OMPHostIRFile.empty())
873     return;
874 
875   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTargetDebug,
876                               "__omp_rtl_debug_kind");
877   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTeamSubscription,
878                               "__omp_rtl_assume_teams_oversubscription");
879   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPThreadSubscription,
880                               "__omp_rtl_assume_threads_oversubscription");
881   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoThreadState,
882                               "__omp_rtl_assume_no_thread_state");
883   OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoNestedParallelism,
884                               "__omp_rtl_assume_no_nested_parallelism");
885 }
886 
887 void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF,
888                                               ProcBindKind ProcBind,
889                                               SourceLocation Loc) {
890   // Do nothing in case of SPMD mode and L0 parallel.
891   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
892     return;
893 
894   CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc);
895 }
896 
897 void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF,
898                                                 llvm::Value *NumThreads,
899                                                 SourceLocation Loc) {
900   // Nothing to do.
901 }
902 
903 void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF,
904                                               const Expr *NumTeams,
905                                               const Expr *ThreadLimit,
906                                               SourceLocation Loc) {}
907 
908 llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction(
909     const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
910     OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
911   // Emit target region as a standalone region.
912   bool PrevIsInTTDRegion = IsInTTDRegion;
913   IsInTTDRegion = false;
914   auto *OutlinedFun =
915       cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction(
916           D, ThreadIDVar, InnermostKind, CodeGen));
917   IsInTTDRegion = PrevIsInTTDRegion;
918   if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) {
919     llvm::Function *WrapperFun =
920         createParallelDataSharingWrapper(OutlinedFun, D);
921     WrapperFunctionsMap[OutlinedFun] = WrapperFun;
922   }
923 
924   return OutlinedFun;
925 }
926 
927 /// Get list of lastprivate variables from the teams distribute ... or
928 /// teams {distribute ...} directives.
929 static void
930 getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D,
931                              llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
932   assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&
933          "expected teams directive.");
934   const OMPExecutableDirective *Dir = &D;
935   if (!isOpenMPDistributeDirective(D.getDirectiveKind())) {
936     if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild(
937             Ctx,
938             D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(
939                 /*IgnoreCaptured=*/true))) {
940       Dir = dyn_cast_or_null<OMPExecutableDirective>(S);
941       if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind()))
942         Dir = nullptr;
943     }
944   }
945   if (!Dir)
946     return;
947   for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) {
948     for (const Expr *E : C->getVarRefs())
949       Vars.push_back(getPrivateItem(E));
950   }
951 }
952 
953 /// Get list of reduction variables from the teams ... directives.
954 static void
955 getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D,
956                       llvm::SmallVectorImpl<const ValueDecl *> &Vars) {
957   assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&
958          "expected teams directive.");
959   for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
960     for (const Expr *E : C->privates())
961       Vars.push_back(getPrivateItem(E));
962   }
963 }
964 
965 llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction(
966     const OMPExecutableDirective &D, const VarDecl *ThreadIDVar,
967     OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) {
968   SourceLocation Loc = D.getBeginLoc();
969 
970   const RecordDecl *GlobalizedRD = nullptr;
971   llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions;
972   llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields;
973   unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size;
974   // Globalize team reductions variable unconditionally in all modes.
975   if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
976     getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions);
977   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
978     getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions);
979     if (!LastPrivatesReductions.empty()) {
980       GlobalizedRD = ::buildRecordForGlobalizedVars(
981           CGM.getContext(), std::nullopt, LastPrivatesReductions,
982           MappedDeclsFields, WarpSize);
983     }
984   } else if (!LastPrivatesReductions.empty()) {
985     assert(!TeamAndReductions.first &&
986            "Previous team declaration is not expected.");
987     TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl();
988     std::swap(TeamAndReductions.second, LastPrivatesReductions);
989   }
990 
991   // Emit target region as a standalone region.
992   class NVPTXPrePostActionTy : public PrePostActionTy {
993     SourceLocation &Loc;
994     const RecordDecl *GlobalizedRD;
995     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
996         &MappedDeclsFields;
997 
998   public:
999     NVPTXPrePostActionTy(
1000         SourceLocation &Loc, const RecordDecl *GlobalizedRD,
1001         llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
1002             &MappedDeclsFields)
1003         : Loc(Loc), GlobalizedRD(GlobalizedRD),
1004           MappedDeclsFields(MappedDeclsFields) {}
1005     void Enter(CodeGenFunction &CGF) override {
1006       auto &Rt =
1007           static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1008       if (GlobalizedRD) {
1009         auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
1010         I->getSecond().MappedParams =
1011             std::make_unique<CodeGenFunction::OMPMapVars>();
1012         DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
1013         for (const auto &Pair : MappedDeclsFields) {
1014           assert(Pair.getFirst()->isCanonicalDecl() &&
1015                  "Expected canonical declaration");
1016           Data.insert(std::make_pair(Pair.getFirst(), MappedVarData()));
1017         }
1018       }
1019       Rt.emitGenericVarsProlog(CGF, Loc);
1020     }
1021     void Exit(CodeGenFunction &CGF) override {
1022       static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
1023           .emitGenericVarsEpilog(CGF);
1024     }
1025   } Action(Loc, GlobalizedRD, MappedDeclsFields);
1026   CodeGen.setAction(Action);
1027   llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction(
1028       D, ThreadIDVar, InnermostKind, CodeGen);
1029 
1030   return OutlinedFun;
1031 }
1032 
1033 void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF,
1034                                                  SourceLocation Loc,
1035                                                  bool WithSPMDCheck) {
1036   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic &&
1037       getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
1038     return;
1039 
1040   CGBuilderTy &Bld = CGF.Builder;
1041 
1042   const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
1043   if (I == FunctionGlobalizedDecls.end())
1044     return;
1045 
1046   for (auto &Rec : I->getSecond().LocalVarData) {
1047     const auto *VD = cast<VarDecl>(Rec.first);
1048     bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first);
1049     QualType VarTy = VD->getType();
1050 
1051     // Get the local allocation of a firstprivate variable before sharing
1052     llvm::Value *ParValue;
1053     if (EscapedParam) {
1054       LValue ParLVal =
1055           CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
1056       ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc);
1057     }
1058 
1059     // Allocate space for the variable to be globalized
1060     llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())};
1061     llvm::CallBase *VoidPtr =
1062         CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1063                                 CGM.getModule(), OMPRTL___kmpc_alloc_shared),
1064                             AllocArgs, VD->getName());
1065     // FIXME: We should use the variables actual alignment as an argument.
1066     VoidPtr->addRetAttr(llvm::Attribute::get(
1067         CGM.getLLVMContext(), llvm::Attribute::Alignment,
1068         CGM.getContext().getTargetInfo().getNewAlign() / 8));
1069 
1070     // Cast the void pointer and get the address of the globalized variable.
1071     llvm::PointerType *VarPtrTy = CGF.ConvertTypeForMem(VarTy)->getPointerTo();
1072     llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
1073         VoidPtr, VarPtrTy, VD->getName() + "_on_stack");
1074     LValue VarAddr = CGF.MakeNaturalAlignAddrLValue(CastedVoidPtr, VarTy);
1075     Rec.second.PrivateAddr = VarAddr.getAddress(CGF);
1076     Rec.second.GlobalizedVal = VoidPtr;
1077 
1078     // Assign the local allocation to the newly globalized location.
1079     if (EscapedParam) {
1080       CGF.EmitStoreOfScalar(ParValue, VarAddr);
1081       I->getSecond().MappedParams->setVarAddr(CGF, VD, VarAddr.getAddress(CGF));
1082     }
1083     if (auto *DI = CGF.getDebugInfo())
1084       VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(VD->getLocation()));
1085   }
1086   for (const auto *VD : I->getSecond().EscapedVariableLengthDecls) {
1087     // Use actual memory size of the VLA object including the padding
1088     // for alignment purposes.
1089     llvm::Value *Size = CGF.getTypeSize(VD->getType());
1090     CharUnits Align = CGM.getContext().getDeclAlign(VD);
1091     Size = Bld.CreateNUWAdd(
1092         Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1));
1093     llvm::Value *AlignVal =
1094         llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity());
1095 
1096     Size = Bld.CreateUDiv(Size, AlignVal);
1097     Size = Bld.CreateNUWMul(Size, AlignVal);
1098 
1099     // Allocate space for this VLA object to be globalized.
1100     llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())};
1101     llvm::CallBase *VoidPtr =
1102         CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1103                                 CGM.getModule(), OMPRTL___kmpc_alloc_shared),
1104                             AllocArgs, VD->getName());
1105     VoidPtr->addRetAttr(
1106         llvm::Attribute::get(CGM.getLLVMContext(), llvm::Attribute::Alignment,
1107                              CGM.getContext().getTargetInfo().getNewAlign()));
1108 
1109     I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(
1110         std::pair<llvm::Value *, llvm::Value *>(
1111             {VoidPtr, CGF.getTypeSize(VD->getType())}));
1112     LValue Base = CGF.MakeAddrLValue(VoidPtr, VD->getType(),
1113                                      CGM.getContext().getDeclAlign(VD),
1114                                      AlignmentSource::Decl);
1115     I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD),
1116                                             Base.getAddress(CGF));
1117   }
1118   I->getSecond().MappedParams->apply(CGF);
1119 }
1120 
1121 void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF,
1122                                                  bool WithSPMDCheck) {
1123   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic &&
1124       getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD)
1125     return;
1126 
1127   const auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
1128   if (I != FunctionGlobalizedDecls.end()) {
1129     // Deallocate the memory for each globalized VLA object
1130     for (auto AddrSizePair :
1131          llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) {
1132       CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1133                               CGM.getModule(), OMPRTL___kmpc_free_shared),
1134                           {AddrSizePair.first, AddrSizePair.second});
1135     }
1136     // Deallocate the memory for each globalized value
1137     for (auto &Rec : llvm::reverse(I->getSecond().LocalVarData)) {
1138       const auto *VD = cast<VarDecl>(Rec.first);
1139       I->getSecond().MappedParams->restore(CGF);
1140 
1141       llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal,
1142                                  CGF.getTypeSize(VD->getType())};
1143       CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1144                               CGM.getModule(), OMPRTL___kmpc_free_shared),
1145                           FreeArgs);
1146     }
1147   }
1148 }
1149 
1150 void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF,
1151                                          const OMPExecutableDirective &D,
1152                                          SourceLocation Loc,
1153                                          llvm::Function *OutlinedFn,
1154                                          ArrayRef<llvm::Value *> CapturedVars) {
1155   if (!CGF.HaveInsertPoint())
1156     return;
1157 
1158   Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
1159                                                       /*Name=*/".zero.addr");
1160   CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr);
1161   llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs;
1162   OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer());
1163   OutlinedFnArgs.push_back(ZeroAddr.getPointer());
1164   OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end());
1165   emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs);
1166 }
1167 
1168 void CGOpenMPRuntimeGPU::emitParallelCall(CodeGenFunction &CGF,
1169                                           SourceLocation Loc,
1170                                           llvm::Function *OutlinedFn,
1171                                           ArrayRef<llvm::Value *> CapturedVars,
1172                                           const Expr *IfCond,
1173                                           llvm::Value *NumThreads) {
1174   if (!CGF.HaveInsertPoint())
1175     return;
1176 
1177   auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond,
1178                         NumThreads](CodeGenFunction &CGF,
1179                                     PrePostActionTy &Action) {
1180     CGBuilderTy &Bld = CGF.Builder;
1181     llvm::Value *NumThreadsVal = NumThreads;
1182     llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn];
1183     llvm::Value *ID = llvm::ConstantPointerNull::get(CGM.Int8PtrTy);
1184     if (WFn)
1185       ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy);
1186     llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(OutlinedFn, CGM.Int8PtrTy);
1187 
1188     // Create a private scope that will globalize the arguments
1189     // passed from the outside of the target region.
1190     // TODO: Is that needed?
1191     CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF);
1192 
1193     Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca(
1194         llvm::ArrayType::get(CGM.VoidPtrTy, CapturedVars.size()),
1195         "captured_vars_addrs");
1196     // There's something to share.
1197     if (!CapturedVars.empty()) {
1198       // Prepare for parallel region. Indicate the outlined function.
1199       ASTContext &Ctx = CGF.getContext();
1200       unsigned Idx = 0;
1201       for (llvm::Value *V : CapturedVars) {
1202         Address Dst = Bld.CreateConstArrayGEP(CapturedVarsAddrs, Idx);
1203         llvm::Value *PtrV;
1204         if (V->getType()->isIntegerTy())
1205           PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy);
1206         else
1207           PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy);
1208         CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false,
1209                               Ctx.getPointerType(Ctx.VoidPtrTy));
1210         ++Idx;
1211       }
1212     }
1213 
1214     llvm::Value *IfCondVal = nullptr;
1215     if (IfCond)
1216       IfCondVal = Bld.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.Int32Ty,
1217                                     /* isSigned */ false);
1218     else
1219       IfCondVal = llvm::ConstantInt::get(CGF.Int32Ty, 1);
1220 
1221     if (!NumThreadsVal)
1222       NumThreadsVal = llvm::ConstantInt::get(CGF.Int32Ty, -1);
1223     else
1224       NumThreadsVal = Bld.CreateZExtOrTrunc(NumThreadsVal, CGF.Int32Ty),
1225 
1226       assert(IfCondVal && "Expected a value");
1227     llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
1228     llvm::Value *Args[] = {
1229         RTLoc,
1230         getThreadID(CGF, Loc),
1231         IfCondVal,
1232         NumThreadsVal,
1233         llvm::ConstantInt::get(CGF.Int32Ty, -1),
1234         FnPtr,
1235         ID,
1236         Bld.CreateBitOrPointerCast(CapturedVarsAddrs.getPointer(),
1237                                    CGF.VoidPtrPtrTy),
1238         llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())};
1239     CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1240                             CGM.getModule(), OMPRTL___kmpc_parallel_51),
1241                         Args);
1242   };
1243 
1244   RegionCodeGenTy RCG(ParallelGen);
1245   RCG(CGF);
1246 }
1247 
1248 void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) {
1249   // Always emit simple barriers!
1250   if (!CGF.HaveInsertPoint())
1251     return;
1252   // Build call __kmpc_barrier_simple_spmd(nullptr, 0);
1253   // This function does not use parameters, so we can emit just default values.
1254   llvm::Value *Args[] = {
1255       llvm::ConstantPointerNull::get(
1256           cast<llvm::PointerType>(getIdentTyPointerTy())),
1257       llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)};
1258   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1259                           CGM.getModule(), OMPRTL___kmpc_barrier_simple_spmd),
1260                       Args);
1261 }
1262 
1263 void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF,
1264                                            SourceLocation Loc,
1265                                            OpenMPDirectiveKind Kind, bool,
1266                                            bool) {
1267   // Always emit simple barriers!
1268   if (!CGF.HaveInsertPoint())
1269     return;
1270   // Build call __kmpc_cancel_barrier(loc, thread_id);
1271   unsigned Flags = getDefaultFlagsForBarriers(Kind);
1272   llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags),
1273                          getThreadID(CGF, Loc)};
1274 
1275   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1276                           CGM.getModule(), OMPRTL___kmpc_barrier),
1277                       Args);
1278 }
1279 
1280 void CGOpenMPRuntimeGPU::emitCriticalRegion(
1281     CodeGenFunction &CGF, StringRef CriticalName,
1282     const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc,
1283     const Expr *Hint) {
1284   llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop");
1285   llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test");
1286   llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync");
1287   llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body");
1288   llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit");
1289 
1290   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1291 
1292   // Get the mask of active threads in the warp.
1293   llvm::Value *Mask = CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1294       CGM.getModule(), OMPRTL___kmpc_warp_active_thread_mask));
1295   // Fetch team-local id of the thread.
1296   llvm::Value *ThreadID = RT.getGPUThreadID(CGF);
1297 
1298   // Get the width of the team.
1299   llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF);
1300 
1301   // Initialize the counter variable for the loop.
1302   QualType Int32Ty =
1303       CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0);
1304   Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter");
1305   LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty);
1306   CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal,
1307                         /*isInit=*/true);
1308 
1309   // Block checks if loop counter exceeds upper bound.
1310   CGF.EmitBlock(LoopBB);
1311   llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
1312   llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth);
1313   CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB);
1314 
1315   // Block tests which single thread should execute region, and which threads
1316   // should go straight to synchronisation point.
1317   CGF.EmitBlock(TestBB);
1318   CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc);
1319   llvm::Value *CmpThreadToCounter =
1320       CGF.Builder.CreateICmpEQ(ThreadID, CounterVal);
1321   CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB);
1322 
1323   // Block emits the body of the critical region.
1324   CGF.EmitBlock(BodyBB);
1325 
1326   // Output the critical statement.
1327   CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc,
1328                                       Hint);
1329 
1330   // After the body surrounded by the critical region, the single executing
1331   // thread will jump to the synchronisation point.
1332   // Block waits for all threads in current team to finish then increments the
1333   // counter variable and returns to the loop.
1334   CGF.EmitBlock(SyncBB);
1335   // Reconverge active threads in the warp.
1336   (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
1337                                 CGM.getModule(), OMPRTL___kmpc_syncwarp),
1338                             Mask);
1339 
1340   llvm::Value *IncCounterVal =
1341       CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1));
1342   CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal);
1343   CGF.EmitBranch(LoopBB);
1344 
1345   // Block that is reached when  all threads in the team complete the region.
1346   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1347 }
1348 
1349 /// Cast value to the specified type.
1350 static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val,
1351                                     QualType ValTy, QualType CastTy,
1352                                     SourceLocation Loc) {
1353   assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() &&
1354          "Cast type must sized.");
1355   assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() &&
1356          "Val type must sized.");
1357   llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy);
1358   if (ValTy == CastTy)
1359     return Val;
1360   if (CGF.getContext().getTypeSizeInChars(ValTy) ==
1361       CGF.getContext().getTypeSizeInChars(CastTy))
1362     return CGF.Builder.CreateBitCast(Val, LLVMCastTy);
1363   if (CastTy->isIntegerType() && ValTy->isIntegerType())
1364     return CGF.Builder.CreateIntCast(Val, LLVMCastTy,
1365                                      CastTy->hasSignedIntegerRepresentation());
1366   Address CastItem = CGF.CreateMemTemp(CastTy);
1367   Address ValCastItem = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
1368       CastItem, Val->getType()->getPointerTo(CastItem.getAddressSpace()),
1369       Val->getType());
1370   CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy,
1371                         LValueBaseInfo(AlignmentSource::Type),
1372                         TBAAAccessInfo());
1373   return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc,
1374                               LValueBaseInfo(AlignmentSource::Type),
1375                               TBAAAccessInfo());
1376 }
1377 
1378 /// This function creates calls to one of two shuffle functions to copy
1379 /// variables between lanes in a warp.
1380 static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF,
1381                                                  llvm::Value *Elem,
1382                                                  QualType ElemType,
1383                                                  llvm::Value *Offset,
1384                                                  SourceLocation Loc) {
1385   CodeGenModule &CGM = CGF.CGM;
1386   CGBuilderTy &Bld = CGF.Builder;
1387   CGOpenMPRuntimeGPU &RT =
1388       *(static_cast<CGOpenMPRuntimeGPU *>(&CGM.getOpenMPRuntime()));
1389   llvm::OpenMPIRBuilder &OMPBuilder = RT.getOMPBuilder();
1390 
1391   CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType);
1392   assert(Size.getQuantity() <= 8 &&
1393          "Unsupported bitwidth in shuffle instruction.");
1394 
1395   RuntimeFunction ShuffleFn = Size.getQuantity() <= 4
1396                                   ? OMPRTL___kmpc_shuffle_int32
1397                                   : OMPRTL___kmpc_shuffle_int64;
1398 
1399   // Cast all types to 32- or 64-bit values before calling shuffle routines.
1400   QualType CastTy = CGF.getContext().getIntTypeForBitwidth(
1401       Size.getQuantity() <= 4 ? 32 : 64, /*Signed=*/1);
1402   llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc);
1403   llvm::Value *WarpSize =
1404       Bld.CreateIntCast(RT.getGPUWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true);
1405 
1406   llvm::Value *ShuffledVal = CGF.EmitRuntimeCall(
1407       OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), ShuffleFn),
1408       {ElemCast, Offset, WarpSize});
1409 
1410   return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc);
1411 }
1412 
1413 static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr,
1414                             Address DestAddr, QualType ElemType,
1415                             llvm::Value *Offset, SourceLocation Loc) {
1416   CGBuilderTy &Bld = CGF.Builder;
1417 
1418   CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType);
1419   // Create the loop over the big sized data.
1420   // ptr = (void*)Elem;
1421   // ptrEnd = (void*) Elem + 1;
1422   // Step = 8;
1423   // while (ptr + Step < ptrEnd)
1424   //   shuffle((int64_t)*ptr);
1425   // Step = 4;
1426   // while (ptr + Step < ptrEnd)
1427   //   shuffle((int32_t)*ptr);
1428   // ...
1429   Address ElemPtr = DestAddr;
1430   Address Ptr = SrcAddr;
1431   Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast(
1432       Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy, CGF.Int8Ty);
1433   for (int IntSize = 8; IntSize >= 1; IntSize /= 2) {
1434     if (Size < CharUnits::fromQuantity(IntSize))
1435       continue;
1436     QualType IntType = CGF.getContext().getIntTypeForBitwidth(
1437         CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)),
1438         /*Signed=*/1);
1439     llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType);
1440     Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo(),
1441                                                   IntTy);
1442     ElemPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
1443         ElemPtr, IntTy->getPointerTo(), IntTy);
1444     if (Size.getQuantity() / IntSize > 1) {
1445       llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond");
1446       llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then");
1447       llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit");
1448       llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock();
1449       CGF.EmitBlock(PreCondBB);
1450       llvm::PHINode *PhiSrc =
1451           Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2);
1452       PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB);
1453       llvm::PHINode *PhiDest =
1454           Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2);
1455       PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB);
1456       Ptr = Address(PhiSrc, Ptr.getElementType(), Ptr.getAlignment());
1457       ElemPtr =
1458           Address(PhiDest, ElemPtr.getElementType(), ElemPtr.getAlignment());
1459       llvm::Value *PtrDiff = Bld.CreatePtrDiff(
1460           CGF.Int8Ty, PtrEnd.getPointer(),
1461           Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr.getPointer(),
1462                                                   CGF.VoidPtrTy));
1463       Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)),
1464                        ThenBB, ExitBB);
1465       CGF.EmitBlock(ThenBB);
1466       llvm::Value *Res = createRuntimeShuffleFunction(
1467           CGF,
1468           CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc,
1469                                LValueBaseInfo(AlignmentSource::Type),
1470                                TBAAAccessInfo()),
1471           IntType, Offset, Loc);
1472       CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType,
1473                             LValueBaseInfo(AlignmentSource::Type),
1474                             TBAAAccessInfo());
1475       Address LocalPtr = Bld.CreateConstGEP(Ptr, 1);
1476       Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1);
1477       PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB);
1478       PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB);
1479       CGF.EmitBranch(PreCondBB);
1480       CGF.EmitBlock(ExitBB);
1481     } else {
1482       llvm::Value *Res = createRuntimeShuffleFunction(
1483           CGF,
1484           CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc,
1485                                LValueBaseInfo(AlignmentSource::Type),
1486                                TBAAAccessInfo()),
1487           IntType, Offset, Loc);
1488       CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType,
1489                             LValueBaseInfo(AlignmentSource::Type),
1490                             TBAAAccessInfo());
1491       Ptr = Bld.CreateConstGEP(Ptr, 1);
1492       ElemPtr = Bld.CreateConstGEP(ElemPtr, 1);
1493     }
1494     Size = Size % IntSize;
1495   }
1496 }
1497 
1498 namespace {
1499 enum CopyAction : unsigned {
1500   // RemoteLaneToThread: Copy over a Reduce list from a remote lane in
1501   // the warp using shuffle instructions.
1502   RemoteLaneToThread,
1503   // ThreadCopy: Make a copy of a Reduce list on the thread's stack.
1504   ThreadCopy,
1505   // ThreadToScratchpad: Copy a team-reduced array to the scratchpad.
1506   ThreadToScratchpad,
1507   // ScratchpadToThread: Copy from a scratchpad array in global memory
1508   // containing team-reduced data to a thread's stack.
1509   ScratchpadToThread,
1510 };
1511 } // namespace
1512 
1513 struct CopyOptionsTy {
1514   llvm::Value *RemoteLaneOffset;
1515   llvm::Value *ScratchpadIndex;
1516   llvm::Value *ScratchpadWidth;
1517 };
1518 
1519 /// Emit instructions to copy a Reduce list, which contains partially
1520 /// aggregated values, in the specified direction.
1521 static void emitReductionListCopy(
1522     CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy,
1523     ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase,
1524     CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) {
1525 
1526   CodeGenModule &CGM = CGF.CGM;
1527   ASTContext &C = CGM.getContext();
1528   CGBuilderTy &Bld = CGF.Builder;
1529 
1530   llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset;
1531   llvm::Value *ScratchpadIndex = CopyOptions.ScratchpadIndex;
1532   llvm::Value *ScratchpadWidth = CopyOptions.ScratchpadWidth;
1533 
1534   // Iterates, element-by-element, through the source Reduce list and
1535   // make a copy.
1536   unsigned Idx = 0;
1537   unsigned Size = Privates.size();
1538   for (const Expr *Private : Privates) {
1539     Address SrcElementAddr = Address::invalid();
1540     Address DestElementAddr = Address::invalid();
1541     Address DestElementPtrAddr = Address::invalid();
1542     // Should we shuffle in an element from a remote lane?
1543     bool ShuffleInElement = false;
1544     // Set to true to update the pointer in the dest Reduce list to a
1545     // newly created element.
1546     bool UpdateDestListPtr = false;
1547     // Increment the src or dest pointer to the scratchpad, for each
1548     // new element.
1549     bool IncrScratchpadSrc = false;
1550     bool IncrScratchpadDest = false;
1551     QualType PrivatePtrType = C.getPointerType(Private->getType());
1552     llvm::Type *PrivateLlvmPtrType = CGF.ConvertType(PrivatePtrType);
1553 
1554     switch (Action) {
1555     case RemoteLaneToThread: {
1556       // Step 1.1: Get the address for the src element in the Reduce list.
1557       Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
1558       SrcElementAddr =
1559           CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast(
1560                                     SrcElementPtrAddr, PrivateLlvmPtrType),
1561                                 PrivatePtrType->castAs<PointerType>());
1562 
1563       // Step 1.2: Create a temporary to store the element in the destination
1564       // Reduce list.
1565       DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
1566       DestElementAddr =
1567           CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element");
1568       ShuffleInElement = true;
1569       UpdateDestListPtr = true;
1570       break;
1571     }
1572     case ThreadCopy: {
1573       // Step 1.1: Get the address for the src element in the Reduce list.
1574       Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
1575       SrcElementAddr =
1576           CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast(
1577                                     SrcElementPtrAddr, PrivateLlvmPtrType),
1578                                 PrivatePtrType->castAs<PointerType>());
1579 
1580       // Step 1.2: Get the address for dest element.  The destination
1581       // element has already been created on the thread's stack.
1582       DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
1583       DestElementAddr =
1584           CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast(
1585                                     DestElementPtrAddr, PrivateLlvmPtrType),
1586                                 PrivatePtrType->castAs<PointerType>());
1587       break;
1588     }
1589     case ThreadToScratchpad: {
1590       // Step 1.1: Get the address for the src element in the Reduce list.
1591       Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx);
1592       SrcElementAddr =
1593           CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast(
1594                                     SrcElementPtrAddr, PrivateLlvmPtrType),
1595                                 PrivatePtrType->castAs<PointerType>());
1596 
1597       // Step 1.2: Get the address for dest element:
1598       // address = base + index * ElementSizeInChars.
1599       llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
1600       llvm::Value *CurrentOffset =
1601           Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex);
1602       llvm::Value *ScratchPadElemAbsolutePtrVal =
1603           Bld.CreateNUWAdd(DestBase.getPointer(), CurrentOffset);
1604       ScratchPadElemAbsolutePtrVal =
1605           Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy);
1606       DestElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty,
1607                                 C.getTypeAlignInChars(Private->getType()));
1608       IncrScratchpadDest = true;
1609       break;
1610     }
1611     case ScratchpadToThread: {
1612       // Step 1.1: Get the address for the src element in the scratchpad.
1613       // address = base + index * ElementSizeInChars.
1614       llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
1615       llvm::Value *CurrentOffset =
1616           Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex);
1617       llvm::Value *ScratchPadElemAbsolutePtrVal =
1618           Bld.CreateNUWAdd(SrcBase.getPointer(), CurrentOffset);
1619       ScratchPadElemAbsolutePtrVal =
1620           Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy);
1621       SrcElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty,
1622                                C.getTypeAlignInChars(Private->getType()));
1623       IncrScratchpadSrc = true;
1624 
1625       // Step 1.2: Create a temporary to store the element in the destination
1626       // Reduce list.
1627       DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx);
1628       DestElementAddr =
1629           CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element");
1630       UpdateDestListPtr = true;
1631       break;
1632     }
1633     }
1634 
1635     // Regardless of src and dest of copy, we emit the load of src
1636     // element as this is required in all directions
1637     SrcElementAddr = Bld.CreateElementBitCast(
1638         SrcElementAddr, CGF.ConvertTypeForMem(Private->getType()));
1639     DestElementAddr = Bld.CreateElementBitCast(DestElementAddr,
1640                                                SrcElementAddr.getElementType());
1641 
1642     // Now that all active lanes have read the element in the
1643     // Reduce list, shuffle over the value from the remote lane.
1644     if (ShuffleInElement) {
1645       shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(),
1646                       RemoteLaneOffset, Private->getExprLoc());
1647     } else {
1648       switch (CGF.getEvaluationKind(Private->getType())) {
1649       case TEK_Scalar: {
1650         llvm::Value *Elem = CGF.EmitLoadOfScalar(
1651             SrcElementAddr, /*Volatile=*/false, Private->getType(),
1652             Private->getExprLoc(), LValueBaseInfo(AlignmentSource::Type),
1653             TBAAAccessInfo());
1654         // Store the source element value to the dest element address.
1655         CGF.EmitStoreOfScalar(
1656             Elem, DestElementAddr, /*Volatile=*/false, Private->getType(),
1657             LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
1658         break;
1659       }
1660       case TEK_Complex: {
1661         CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex(
1662             CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
1663             Private->getExprLoc());
1664         CGF.EmitStoreOfComplex(
1665             Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
1666             /*isInit=*/false);
1667         break;
1668       }
1669       case TEK_Aggregate:
1670         CGF.EmitAggregateCopy(
1671             CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
1672             CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
1673             Private->getType(), AggValueSlot::DoesNotOverlap);
1674         break;
1675       }
1676     }
1677 
1678     // Step 3.1: Modify reference in dest Reduce list as needed.
1679     // Modifying the reference in Reduce list to point to the newly
1680     // created element.  The element is live in the current function
1681     // scope and that of functions it invokes (i.e., reduce_function).
1682     // RemoteReduceData[i] = (void*)&RemoteElem
1683     if (UpdateDestListPtr) {
1684       CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast(
1685                                 DestElementAddr.getPointer(), CGF.VoidPtrTy),
1686                             DestElementPtrAddr, /*Volatile=*/false,
1687                             C.VoidPtrTy);
1688     }
1689 
1690     // Step 4.1: Increment SrcBase/DestBase so that it points to the starting
1691     // address of the next element in scratchpad memory, unless we're currently
1692     // processing the last one.  Memory alignment is also taken care of here.
1693     if ((IncrScratchpadDest || IncrScratchpadSrc) && (Idx + 1 < Size)) {
1694       // FIXME: This code doesn't make any sense, it's trying to perform
1695       // integer arithmetic on pointers.
1696       llvm::Value *ScratchpadBasePtr =
1697           IncrScratchpadDest ? DestBase.getPointer() : SrcBase.getPointer();
1698       llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType());
1699       ScratchpadBasePtr = Bld.CreateNUWAdd(
1700           ScratchpadBasePtr,
1701           Bld.CreateNUWMul(ScratchpadWidth, ElementSizeInChars));
1702 
1703       // Take care of global memory alignment for performance
1704       ScratchpadBasePtr = Bld.CreateNUWSub(
1705           ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1));
1706       ScratchpadBasePtr = Bld.CreateUDiv(
1707           ScratchpadBasePtr,
1708           llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment));
1709       ScratchpadBasePtr = Bld.CreateNUWAdd(
1710           ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1));
1711       ScratchpadBasePtr = Bld.CreateNUWMul(
1712           ScratchpadBasePtr,
1713           llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment));
1714 
1715       if (IncrScratchpadDest)
1716         DestBase =
1717             Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign());
1718       else /* IncrScratchpadSrc = true */
1719         SrcBase =
1720             Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign());
1721     }
1722 
1723     ++Idx;
1724   }
1725 }
1726 
1727 /// This function emits a helper that gathers Reduce lists from the first
1728 /// lane of every active warp to lanes in the first warp.
1729 ///
1730 /// void inter_warp_copy_func(void* reduce_data, num_warps)
1731 ///   shared smem[warp_size];
1732 ///   For all data entries D in reduce_data:
1733 ///     sync
1734 ///     If (I am the first lane in each warp)
1735 ///       Copy my local D to smem[warp_id]
1736 ///     sync
1737 ///     if (I am the first warp)
1738 ///       Copy smem[thread_id] to my local D
1739 static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM,
1740                                               ArrayRef<const Expr *> Privates,
1741                                               QualType ReductionArrayTy,
1742                                               SourceLocation Loc) {
1743   ASTContext &C = CGM.getContext();
1744   llvm::Module &M = CGM.getModule();
1745 
1746   // ReduceList: thread local Reduce list.
1747   // At the stage of the computation when this function is called, partially
1748   // aggregated values reside in the first lane of every active warp.
1749   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
1750                                   C.VoidPtrTy, ImplicitParamDecl::Other);
1751   // NumWarps: number of warps active in the parallel region.  This could
1752   // be smaller than 32 (max warps in a CTA) for partial block reduction.
1753   ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
1754                                 C.getIntTypeForBitwidth(32, /* Signed */ true),
1755                                 ImplicitParamDecl::Other);
1756   FunctionArgList Args;
1757   Args.push_back(&ReduceListArg);
1758   Args.push_back(&NumWarpsArg);
1759 
1760   const CGFunctionInfo &CGFI =
1761       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
1762   auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI),
1763                                     llvm::GlobalValue::InternalLinkage,
1764                                     "_omp_reduction_inter_warp_copy_func", &M);
1765   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
1766   Fn->setDoesNotRecurse();
1767   CodeGenFunction CGF(CGM);
1768   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
1769 
1770   CGBuilderTy &Bld = CGF.Builder;
1771 
1772   // This array is used as a medium to transfer, one reduce element at a time,
1773   // the data from the first lane of every warp to lanes in the first warp
1774   // in order to perform the final step of a reduction in a parallel region
1775   // (reduction across warps).  The array is placed in NVPTX __shared__ memory
1776   // for reduced latency, as well as to have a distinct copy for concurrently
1777   // executing target regions.  The array is declared with common linkage so
1778   // as to be shared across compilation units.
1779   StringRef TransferMediumName =
1780       "__openmp_nvptx_data_transfer_temporary_storage";
1781   llvm::GlobalVariable *TransferMedium =
1782       M.getGlobalVariable(TransferMediumName);
1783   unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size;
1784   if (!TransferMedium) {
1785     auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize);
1786     unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared);
1787     TransferMedium = new llvm::GlobalVariable(
1788         M, Ty, /*isConstant=*/false, llvm::GlobalVariable::WeakAnyLinkage,
1789         llvm::UndefValue::get(Ty), TransferMediumName,
1790         /*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal,
1791         SharedAddressSpace);
1792     CGM.addCompilerUsedGlobal(TransferMedium);
1793   }
1794 
1795   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
1796   // Get the CUDA thread id of the current OpenMP thread on the GPU.
1797   llvm::Value *ThreadID = RT.getGPUThreadID(CGF);
1798   // nvptx_lane_id = nvptx_id % warpsize
1799   llvm::Value *LaneID = getNVPTXLaneID(CGF);
1800   // nvptx_warp_id = nvptx_id / warpsize
1801   llvm::Value *WarpID = getNVPTXWarpID(CGF);
1802 
1803   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
1804   llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
1805   Address LocalReduceList(
1806       Bld.CreatePointerBitCastOrAddrSpaceCast(
1807           CGF.EmitLoadOfScalar(
1808               AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc,
1809               LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()),
1810           ElemTy->getPointerTo()),
1811       ElemTy, CGF.getPointerAlign());
1812 
1813   unsigned Idx = 0;
1814   for (const Expr *Private : Privates) {
1815     //
1816     // Warp master copies reduce element to transfer medium in __shared__
1817     // memory.
1818     //
1819     unsigned RealTySize =
1820         C.getTypeSizeInChars(Private->getType())
1821             .alignTo(C.getTypeAlignInChars(Private->getType()))
1822             .getQuantity();
1823     for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /=2) {
1824       unsigned NumIters = RealTySize / TySize;
1825       if (NumIters == 0)
1826         continue;
1827       QualType CType = C.getIntTypeForBitwidth(
1828           C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1);
1829       llvm::Type *CopyType = CGF.ConvertTypeForMem(CType);
1830       CharUnits Align = CharUnits::fromQuantity(TySize);
1831       llvm::Value *Cnt = nullptr;
1832       Address CntAddr = Address::invalid();
1833       llvm::BasicBlock *PrecondBB = nullptr;
1834       llvm::BasicBlock *ExitBB = nullptr;
1835       if (NumIters > 1) {
1836         CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr");
1837         CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr,
1838                               /*Volatile=*/false, C.IntTy);
1839         PrecondBB = CGF.createBasicBlock("precond");
1840         ExitBB = CGF.createBasicBlock("exit");
1841         llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body");
1842         // There is no need to emit line number for unconditional branch.
1843         (void)ApplyDebugLocation::CreateEmpty(CGF);
1844         CGF.EmitBlock(PrecondBB);
1845         Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc);
1846         llvm::Value *Cmp =
1847             Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters));
1848         Bld.CreateCondBr(Cmp, BodyBB, ExitBB);
1849         CGF.EmitBlock(BodyBB);
1850       }
1851       // kmpc_barrier.
1852       CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
1853                                              /*EmitChecks=*/false,
1854                                              /*ForceSimpleCall=*/true);
1855       llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then");
1856       llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else");
1857       llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont");
1858 
1859       // if (lane_id == 0)
1860       llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master");
1861       Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB);
1862       CGF.EmitBlock(ThenBB);
1863 
1864       // Reduce element = LocalReduceList[i]
1865       Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
1866       llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
1867           ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
1868       // elemptr = ((CopyType*)(elemptrptr)) + I
1869       Address ElemPtr(ElemPtrPtr, CGF.Int8Ty, Align);
1870       ElemPtr = Bld.CreateElementBitCast(ElemPtr, CopyType);
1871       if (NumIters > 1)
1872         ElemPtr = Bld.CreateGEP(ElemPtr, Cnt);
1873 
1874       // Get pointer to location in transfer medium.
1875       // MediumPtr = &medium[warp_id]
1876       llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP(
1877           TransferMedium->getValueType(), TransferMedium,
1878           {llvm::Constant::getNullValue(CGM.Int64Ty), WarpID});
1879       // Casting to actual data type.
1880       // MediumPtr = (CopyType*)MediumPtrAddr;
1881       Address MediumPtr(
1882           Bld.CreateBitCast(
1883               MediumPtrVal,
1884               CopyType->getPointerTo(
1885                   MediumPtrVal->getType()->getPointerAddressSpace())),
1886           CopyType, Align);
1887 
1888       // elem = *elemptr
1889       //*MediumPtr = elem
1890       llvm::Value *Elem = CGF.EmitLoadOfScalar(
1891           ElemPtr, /*Volatile=*/false, CType, Loc,
1892           LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
1893       // Store the source element value to the dest element address.
1894       CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType,
1895                             LValueBaseInfo(AlignmentSource::Type),
1896                             TBAAAccessInfo());
1897 
1898       Bld.CreateBr(MergeBB);
1899 
1900       CGF.EmitBlock(ElseBB);
1901       Bld.CreateBr(MergeBB);
1902 
1903       CGF.EmitBlock(MergeBB);
1904 
1905       // kmpc_barrier.
1906       CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown,
1907                                              /*EmitChecks=*/false,
1908                                              /*ForceSimpleCall=*/true);
1909 
1910       //
1911       // Warp 0 copies reduce element from transfer medium.
1912       //
1913       llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then");
1914       llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else");
1915       llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont");
1916 
1917       Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg);
1918       llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar(
1919           AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc);
1920 
1921       // Up to 32 threads in warp 0 are active.
1922       llvm::Value *IsActiveThread =
1923           Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread");
1924       Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB);
1925 
1926       CGF.EmitBlock(W0ThenBB);
1927 
1928       // SrcMediumPtr = &medium[tid]
1929       llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP(
1930           TransferMedium->getValueType(), TransferMedium,
1931           {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID});
1932       // SrcMediumVal = *SrcMediumPtr;
1933       Address SrcMediumPtr(
1934           Bld.CreateBitCast(
1935               SrcMediumPtrVal,
1936               CopyType->getPointerTo(
1937                   SrcMediumPtrVal->getType()->getPointerAddressSpace())),
1938           CopyType, Align);
1939 
1940       // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I
1941       Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
1942       llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar(
1943           TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc);
1944       Address TargetElemPtr(TargetElemPtrVal, CGF.Int8Ty, Align);
1945       TargetElemPtr = Bld.CreateElementBitCast(TargetElemPtr, CopyType);
1946       if (NumIters > 1)
1947         TargetElemPtr = Bld.CreateGEP(TargetElemPtr, Cnt);
1948 
1949       // *TargetElemPtr = SrcMediumVal;
1950       llvm::Value *SrcMediumValue =
1951           CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc);
1952       CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false,
1953                             CType);
1954       Bld.CreateBr(W0MergeBB);
1955 
1956       CGF.EmitBlock(W0ElseBB);
1957       Bld.CreateBr(W0MergeBB);
1958 
1959       CGF.EmitBlock(W0MergeBB);
1960 
1961       if (NumIters > 1) {
1962         Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1));
1963         CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy);
1964         CGF.EmitBranch(PrecondBB);
1965         (void)ApplyDebugLocation::CreateEmpty(CGF);
1966         CGF.EmitBlock(ExitBB);
1967       }
1968       RealTySize %= TySize;
1969     }
1970     ++Idx;
1971   }
1972 
1973   CGF.FinishFunction();
1974   return Fn;
1975 }
1976 
1977 /// Emit a helper that reduces data across two OpenMP threads (lanes)
1978 /// in the same warp.  It uses shuffle instructions to copy over data from
1979 /// a remote lane's stack.  The reduction algorithm performed is specified
1980 /// by the fourth parameter.
1981 ///
1982 /// Algorithm Versions.
1983 /// Full Warp Reduce (argument value 0):
1984 ///   This algorithm assumes that all 32 lanes are active and gathers
1985 ///   data from these 32 lanes, producing a single resultant value.
1986 /// Contiguous Partial Warp Reduce (argument value 1):
1987 ///   This algorithm assumes that only a *contiguous* subset of lanes
1988 ///   are active.  This happens for the last warp in a parallel region
1989 ///   when the user specified num_threads is not an integer multiple of
1990 ///   32.  This contiguous subset always starts with the zeroth lane.
1991 /// Partial Warp Reduce (argument value 2):
1992 ///   This algorithm gathers data from any number of lanes at any position.
1993 /// All reduced values are stored in the lowest possible lane.  The set
1994 /// of problems every algorithm addresses is a super set of those
1995 /// addressable by algorithms with a lower version number.  Overhead
1996 /// increases as algorithm version increases.
1997 ///
1998 /// Terminology
1999 /// Reduce element:
2000 ///   Reduce element refers to the individual data field with primitive
2001 ///   data types to be combined and reduced across threads.
2002 /// Reduce list:
2003 ///   Reduce list refers to a collection of local, thread-private
2004 ///   reduce elements.
2005 /// Remote Reduce list:
2006 ///   Remote Reduce list refers to a collection of remote (relative to
2007 ///   the current thread) reduce elements.
2008 ///
2009 /// We distinguish between three states of threads that are important to
2010 /// the implementation of this function.
2011 /// Alive threads:
2012 ///   Threads in a warp executing the SIMT instruction, as distinguished from
2013 ///   threads that are inactive due to divergent control flow.
2014 /// Active threads:
2015 ///   The minimal set of threads that has to be alive upon entry to this
2016 ///   function.  The computation is correct iff active threads are alive.
2017 ///   Some threads are alive but they are not active because they do not
2018 ///   contribute to the computation in any useful manner.  Turning them off
2019 ///   may introduce control flow overheads without any tangible benefits.
2020 /// Effective threads:
2021 ///   In order to comply with the argument requirements of the shuffle
2022 ///   function, we must keep all lanes holding data alive.  But at most
2023 ///   half of them perform value aggregation; we refer to this half of
2024 ///   threads as effective. The other half is simply handing off their
2025 ///   data.
2026 ///
2027 /// Procedure
2028 /// Value shuffle:
2029 ///   In this step active threads transfer data from higher lane positions
2030 ///   in the warp to lower lane positions, creating Remote Reduce list.
2031 /// Value aggregation:
2032 ///   In this step, effective threads combine their thread local Reduce list
2033 ///   with Remote Reduce list and store the result in the thread local
2034 ///   Reduce list.
2035 /// Value copy:
2036 ///   In this step, we deal with the assumption made by algorithm 2
2037 ///   (i.e. contiguity assumption).  When we have an odd number of lanes
2038 ///   active, say 2k+1, only k threads will be effective and therefore k
2039 ///   new values will be produced.  However, the Reduce list owned by the
2040 ///   (2k+1)th thread is ignored in the value aggregation.  Therefore
2041 ///   we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so
2042 ///   that the contiguity assumption still holds.
2043 static llvm::Function *emitShuffleAndReduceFunction(
2044     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2045     QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) {
2046   ASTContext &C = CGM.getContext();
2047 
2048   // Thread local Reduce list used to host the values of data to be reduced.
2049   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2050                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2051   // Current lane id; could be logical.
2052   ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy,
2053                               ImplicitParamDecl::Other);
2054   // Offset of the remote source lane relative to the current lane.
2055   ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2056                                         C.ShortTy, ImplicitParamDecl::Other);
2057   // Algorithm version.  This is expected to be known at compile time.
2058   ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2059                                C.ShortTy, ImplicitParamDecl::Other);
2060   FunctionArgList Args;
2061   Args.push_back(&ReduceListArg);
2062   Args.push_back(&LaneIDArg);
2063   Args.push_back(&RemoteLaneOffsetArg);
2064   Args.push_back(&AlgoVerArg);
2065 
2066   const CGFunctionInfo &CGFI =
2067       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2068   auto *Fn = llvm::Function::Create(
2069       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2070       "_omp_reduction_shuffle_and_reduce_func", &CGM.getModule());
2071   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2072   Fn->setDoesNotRecurse();
2073 
2074   CodeGenFunction CGF(CGM);
2075   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2076 
2077   CGBuilderTy &Bld = CGF.Builder;
2078 
2079   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2080   llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
2081   Address LocalReduceList(
2082       Bld.CreatePointerBitCastOrAddrSpaceCast(
2083           CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
2084                                C.VoidPtrTy, SourceLocation()),
2085           ElemTy->getPointerTo()),
2086       ElemTy, CGF.getPointerAlign());
2087 
2088   Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg);
2089   llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar(
2090       AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
2091 
2092   Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg);
2093   llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar(
2094       AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
2095 
2096   Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg);
2097   llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar(
2098       AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation());
2099 
2100   // Create a local thread-private variable to host the Reduce list
2101   // from a remote lane.
2102   Address RemoteReduceList =
2103       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list");
2104 
2105   // This loop iterates through the list of reduce elements and copies,
2106   // element by element, from a remote lane in the warp to RemoteReduceList,
2107   // hosted on the thread's stack.
2108   emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates,
2109                         LocalReduceList, RemoteReduceList,
2110                         {/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal,
2111                          /*ScratchpadIndex=*/nullptr,
2112                          /*ScratchpadWidth=*/nullptr});
2113 
2114   // The actions to be performed on the Remote Reduce list is dependent
2115   // on the algorithm version.
2116   //
2117   //  if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 &&
2118   //  LaneId % 2 == 0 && Offset > 0):
2119   //    do the reduction value aggregation
2120   //
2121   //  The thread local variable Reduce list is mutated in place to host the
2122   //  reduced data, which is the aggregated value produced from local and
2123   //  remote lanes.
2124   //
2125   //  Note that AlgoVer is expected to be a constant integer known at compile
2126   //  time.
2127   //  When AlgoVer==0, the first conjunction evaluates to true, making
2128   //    the entire predicate true during compile time.
2129   //  When AlgoVer==1, the second conjunction has only the second part to be
2130   //    evaluated during runtime.  Other conjunctions evaluates to false
2131   //    during compile time.
2132   //  When AlgoVer==2, the third conjunction has only the second part to be
2133   //    evaluated during runtime.  Other conjunctions evaluates to false
2134   //    during compile time.
2135   llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal);
2136 
2137   llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1));
2138   llvm::Value *CondAlgo1 = Bld.CreateAnd(
2139       Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal));
2140 
2141   llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2));
2142   llvm::Value *CondAlgo2 = Bld.CreateAnd(
2143       Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1))));
2144   CondAlgo2 = Bld.CreateAnd(
2145       CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0)));
2146 
2147   llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1);
2148   CondReduce = Bld.CreateOr(CondReduce, CondAlgo2);
2149 
2150   llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then");
2151   llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else");
2152   llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont");
2153   Bld.CreateCondBr(CondReduce, ThenBB, ElseBB);
2154 
2155   CGF.EmitBlock(ThenBB);
2156   // reduce_function(LocalReduceList, RemoteReduceList)
2157   llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2158       LocalReduceList.getPointer(), CGF.VoidPtrTy);
2159   llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2160       RemoteReduceList.getPointer(), CGF.VoidPtrTy);
2161   CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
2162       CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr});
2163   Bld.CreateBr(MergeBB);
2164 
2165   CGF.EmitBlock(ElseBB);
2166   Bld.CreateBr(MergeBB);
2167 
2168   CGF.EmitBlock(MergeBB);
2169 
2170   // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local
2171   // Reduce list.
2172   Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1));
2173   llvm::Value *CondCopy = Bld.CreateAnd(
2174       Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal));
2175 
2176   llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then");
2177   llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else");
2178   llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont");
2179   Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB);
2180 
2181   CGF.EmitBlock(CpyThenBB);
2182   emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates,
2183                         RemoteReduceList, LocalReduceList);
2184   Bld.CreateBr(CpyMergeBB);
2185 
2186   CGF.EmitBlock(CpyElseBB);
2187   Bld.CreateBr(CpyMergeBB);
2188 
2189   CGF.EmitBlock(CpyMergeBB);
2190 
2191   CGF.FinishFunction();
2192   return Fn;
2193 }
2194 
2195 /// This function emits a helper that copies all the reduction variables from
2196 /// the team into the provided global buffer for the reduction variables.
2197 ///
2198 /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
2199 ///   For all data entries D in reduce_data:
2200 ///     Copy local D to buffer.D[Idx]
2201 static llvm::Value *emitListToGlobalCopyFunction(
2202     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2203     QualType ReductionArrayTy, SourceLocation Loc,
2204     const RecordDecl *TeamReductionRec,
2205     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2206         &VarFieldMap) {
2207   ASTContext &C = CGM.getContext();
2208 
2209   // Buffer: global reduction buffer.
2210   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2211                               C.VoidPtrTy, ImplicitParamDecl::Other);
2212   // Idx: index of the buffer.
2213   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2214                            ImplicitParamDecl::Other);
2215   // ReduceList: thread local Reduce list.
2216   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2217                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2218   FunctionArgList Args;
2219   Args.push_back(&BufferArg);
2220   Args.push_back(&IdxArg);
2221   Args.push_back(&ReduceListArg);
2222 
2223   const CGFunctionInfo &CGFI =
2224       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2225   auto *Fn = llvm::Function::Create(
2226       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2227       "_omp_reduction_list_to_global_copy_func", &CGM.getModule());
2228   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2229   Fn->setDoesNotRecurse();
2230   CodeGenFunction CGF(CGM);
2231   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2232 
2233   CGBuilderTy &Bld = CGF.Builder;
2234 
2235   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2236   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2237   llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
2238   Address LocalReduceList(
2239       Bld.CreatePointerBitCastOrAddrSpaceCast(
2240           CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
2241                                C.VoidPtrTy, Loc),
2242           ElemTy->getPointerTo()),
2243       ElemTy, CGF.getPointerAlign());
2244   QualType StaticTy = C.getRecordType(TeamReductionRec);
2245   llvm::Type *LLVMReductionsBufferTy =
2246       CGM.getTypes().ConvertTypeForMem(StaticTy);
2247   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2248       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2249       LLVMReductionsBufferTy->getPointerTo());
2250   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2251                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2252                                               /*Volatile=*/false, C.IntTy,
2253                                               Loc)};
2254   unsigned Idx = 0;
2255   for (const Expr *Private : Privates) {
2256     // Reduce element = LocalReduceList[i]
2257     Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
2258     llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
2259         ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
2260     // elemptr = ((CopyType*)(elemptrptr)) + I
2261     ElemTy = CGF.ConvertTypeForMem(Private->getType());
2262     ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2263         ElemPtrPtr, ElemTy->getPointerTo());
2264     Address ElemPtr =
2265         Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType()));
2266     const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl();
2267     // Global = Buffer.VD[Idx];
2268     const FieldDecl *FD = VarFieldMap.lookup(VD);
2269     LValue GlobLVal = CGF.EmitLValueForField(
2270         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2271     Address GlobAddr = GlobLVal.getAddress(CGF);
2272     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(),
2273                                                    GlobAddr.getPointer(), Idxs);
2274     GlobLVal.setAddress(Address(BufferPtr,
2275                                 CGF.ConvertTypeForMem(Private->getType()),
2276                                 GlobAddr.getAlignment()));
2277     switch (CGF.getEvaluationKind(Private->getType())) {
2278     case TEK_Scalar: {
2279       llvm::Value *V = CGF.EmitLoadOfScalar(
2280           ElemPtr, /*Volatile=*/false, Private->getType(), Loc,
2281           LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo());
2282       CGF.EmitStoreOfScalar(V, GlobLVal);
2283       break;
2284     }
2285     case TEK_Complex: {
2286       CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(
2287           CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc);
2288       CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false);
2289       break;
2290     }
2291     case TEK_Aggregate:
2292       CGF.EmitAggregateCopy(GlobLVal,
2293                             CGF.MakeAddrLValue(ElemPtr, Private->getType()),
2294                             Private->getType(), AggValueSlot::DoesNotOverlap);
2295       break;
2296     }
2297     ++Idx;
2298   }
2299 
2300   CGF.FinishFunction();
2301   return Fn;
2302 }
2303 
2304 /// This function emits a helper that reduces all the reduction variables from
2305 /// the team into the provided global buffer for the reduction variables.
2306 ///
2307 /// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data)
2308 ///  void *GlobPtrs[];
2309 ///  GlobPtrs[0] = (void*)&buffer.D0[Idx];
2310 ///  ...
2311 ///  GlobPtrs[N] = (void*)&buffer.DN[Idx];
2312 ///  reduce_function(GlobPtrs, reduce_data);
2313 static llvm::Value *emitListToGlobalReduceFunction(
2314     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2315     QualType ReductionArrayTy, SourceLocation Loc,
2316     const RecordDecl *TeamReductionRec,
2317     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2318         &VarFieldMap,
2319     llvm::Function *ReduceFn) {
2320   ASTContext &C = CGM.getContext();
2321 
2322   // Buffer: global reduction buffer.
2323   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2324                               C.VoidPtrTy, ImplicitParamDecl::Other);
2325   // Idx: index of the buffer.
2326   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2327                            ImplicitParamDecl::Other);
2328   // ReduceList: thread local Reduce list.
2329   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2330                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2331   FunctionArgList Args;
2332   Args.push_back(&BufferArg);
2333   Args.push_back(&IdxArg);
2334   Args.push_back(&ReduceListArg);
2335 
2336   const CGFunctionInfo &CGFI =
2337       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2338   auto *Fn = llvm::Function::Create(
2339       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2340       "_omp_reduction_list_to_global_reduce_func", &CGM.getModule());
2341   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2342   Fn->setDoesNotRecurse();
2343   CodeGenFunction CGF(CGM);
2344   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2345 
2346   CGBuilderTy &Bld = CGF.Builder;
2347 
2348   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2349   QualType StaticTy = C.getRecordType(TeamReductionRec);
2350   llvm::Type *LLVMReductionsBufferTy =
2351       CGM.getTypes().ConvertTypeForMem(StaticTy);
2352   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2353       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2354       LLVMReductionsBufferTy->getPointerTo());
2355 
2356   // 1. Build a list of reduction variables.
2357   // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
2358   Address ReductionList =
2359       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
2360   auto IPriv = Privates.begin();
2361   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2362                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2363                                               /*Volatile=*/false, C.IntTy,
2364                                               Loc)};
2365   unsigned Idx = 0;
2366   for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) {
2367     Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2368     // Global = Buffer.VD[Idx];
2369     const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl();
2370     const FieldDecl *FD = VarFieldMap.lookup(VD);
2371     LValue GlobLVal = CGF.EmitLValueForField(
2372         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2373     Address GlobAddr = GlobLVal.getAddress(CGF);
2374     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(
2375         GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs);
2376     llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr);
2377     CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy);
2378     if ((*IPriv)->getType()->isVariablyModifiedType()) {
2379       // Store array size.
2380       ++Idx;
2381       Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2382       llvm::Value *Size = CGF.Builder.CreateIntCast(
2383           CGF.getVLASize(
2384                  CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
2385               .NumElts,
2386           CGF.SizeTy, /*isSigned=*/false);
2387       CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
2388                               Elem);
2389     }
2390   }
2391 
2392   // Call reduce_function(GlobalReduceList, ReduceList)
2393   llvm::Value *GlobalReduceList =
2394       CGF.EmitCastToVoidPtr(ReductionList.getPointer());
2395   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2396   llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar(
2397       AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc);
2398   CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
2399       CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr});
2400   CGF.FinishFunction();
2401   return Fn;
2402 }
2403 
2404 /// This function emits a helper that copies all the reduction variables from
2405 /// the team into the provided global buffer for the reduction variables.
2406 ///
2407 /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data)
2408 ///   For all data entries D in reduce_data:
2409 ///     Copy buffer.D[Idx] to local D;
2410 static llvm::Value *emitGlobalToListCopyFunction(
2411     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2412     QualType ReductionArrayTy, SourceLocation Loc,
2413     const RecordDecl *TeamReductionRec,
2414     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2415         &VarFieldMap) {
2416   ASTContext &C = CGM.getContext();
2417 
2418   // Buffer: global reduction buffer.
2419   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2420                               C.VoidPtrTy, ImplicitParamDecl::Other);
2421   // Idx: index of the buffer.
2422   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2423                            ImplicitParamDecl::Other);
2424   // ReduceList: thread local Reduce list.
2425   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2426                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2427   FunctionArgList Args;
2428   Args.push_back(&BufferArg);
2429   Args.push_back(&IdxArg);
2430   Args.push_back(&ReduceListArg);
2431 
2432   const CGFunctionInfo &CGFI =
2433       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2434   auto *Fn = llvm::Function::Create(
2435       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2436       "_omp_reduction_global_to_list_copy_func", &CGM.getModule());
2437   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2438   Fn->setDoesNotRecurse();
2439   CodeGenFunction CGF(CGM);
2440   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2441 
2442   CGBuilderTy &Bld = CGF.Builder;
2443 
2444   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2445   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2446   llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy);
2447   Address LocalReduceList(
2448       Bld.CreatePointerBitCastOrAddrSpaceCast(
2449           CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false,
2450                                C.VoidPtrTy, Loc),
2451           ElemTy->getPointerTo()),
2452       ElemTy, CGF.getPointerAlign());
2453   QualType StaticTy = C.getRecordType(TeamReductionRec);
2454   llvm::Type *LLVMReductionsBufferTy =
2455       CGM.getTypes().ConvertTypeForMem(StaticTy);
2456   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2457       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2458       LLVMReductionsBufferTy->getPointerTo());
2459 
2460   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2461                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2462                                               /*Volatile=*/false, C.IntTy,
2463                                               Loc)};
2464   unsigned Idx = 0;
2465   for (const Expr *Private : Privates) {
2466     // Reduce element = LocalReduceList[i]
2467     Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx);
2468     llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar(
2469         ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation());
2470     // elemptr = ((CopyType*)(elemptrptr)) + I
2471     ElemTy = CGF.ConvertTypeForMem(Private->getType());
2472     ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2473         ElemPtrPtr, ElemTy->getPointerTo());
2474     Address ElemPtr =
2475         Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType()));
2476     const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl();
2477     // Global = Buffer.VD[Idx];
2478     const FieldDecl *FD = VarFieldMap.lookup(VD);
2479     LValue GlobLVal = CGF.EmitLValueForField(
2480         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2481     Address GlobAddr = GlobLVal.getAddress(CGF);
2482     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(),
2483                                                    GlobAddr.getPointer(), Idxs);
2484     GlobLVal.setAddress(Address(BufferPtr,
2485                                 CGF.ConvertTypeForMem(Private->getType()),
2486                                 GlobAddr.getAlignment()));
2487     switch (CGF.getEvaluationKind(Private->getType())) {
2488     case TEK_Scalar: {
2489       llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc);
2490       CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType(),
2491                             LValueBaseInfo(AlignmentSource::Type),
2492                             TBAAAccessInfo());
2493       break;
2494     }
2495     case TEK_Complex: {
2496       CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc);
2497       CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()),
2498                              /*isInit=*/false);
2499       break;
2500     }
2501     case TEK_Aggregate:
2502       CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()),
2503                             GlobLVal, Private->getType(),
2504                             AggValueSlot::DoesNotOverlap);
2505       break;
2506     }
2507     ++Idx;
2508   }
2509 
2510   CGF.FinishFunction();
2511   return Fn;
2512 }
2513 
2514 /// This function emits a helper that reduces all the reduction variables from
2515 /// the team into the provided global buffer for the reduction variables.
2516 ///
2517 /// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data)
2518 ///  void *GlobPtrs[];
2519 ///  GlobPtrs[0] = (void*)&buffer.D0[Idx];
2520 ///  ...
2521 ///  GlobPtrs[N] = (void*)&buffer.DN[Idx];
2522 ///  reduce_function(reduce_data, GlobPtrs);
2523 static llvm::Value *emitGlobalToListReduceFunction(
2524     CodeGenModule &CGM, ArrayRef<const Expr *> Privates,
2525     QualType ReductionArrayTy, SourceLocation Loc,
2526     const RecordDecl *TeamReductionRec,
2527     const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
2528         &VarFieldMap,
2529     llvm::Function *ReduceFn) {
2530   ASTContext &C = CGM.getContext();
2531 
2532   // Buffer: global reduction buffer.
2533   ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2534                               C.VoidPtrTy, ImplicitParamDecl::Other);
2535   // Idx: index of the buffer.
2536   ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy,
2537                            ImplicitParamDecl::Other);
2538   // ReduceList: thread local Reduce list.
2539   ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr,
2540                                   C.VoidPtrTy, ImplicitParamDecl::Other);
2541   FunctionArgList Args;
2542   Args.push_back(&BufferArg);
2543   Args.push_back(&IdxArg);
2544   Args.push_back(&ReduceListArg);
2545 
2546   const CGFunctionInfo &CGFI =
2547       CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args);
2548   auto *Fn = llvm::Function::Create(
2549       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
2550       "_omp_reduction_global_to_list_reduce_func", &CGM.getModule());
2551   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
2552   Fn->setDoesNotRecurse();
2553   CodeGenFunction CGF(CGM);
2554   CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc);
2555 
2556   CGBuilderTy &Bld = CGF.Builder;
2557 
2558   Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg);
2559   QualType StaticTy = C.getRecordType(TeamReductionRec);
2560   llvm::Type *LLVMReductionsBufferTy =
2561       CGM.getTypes().ConvertTypeForMem(StaticTy);
2562   llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast(
2563       CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc),
2564       LLVMReductionsBufferTy->getPointerTo());
2565 
2566   // 1. Build a list of reduction variables.
2567   // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
2568   Address ReductionList =
2569       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
2570   auto IPriv = Privates.begin();
2571   llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty),
2572                          CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg),
2573                                               /*Volatile=*/false, C.IntTy,
2574                                               Loc)};
2575   unsigned Idx = 0;
2576   for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) {
2577     Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2578     // Global = Buffer.VD[Idx];
2579     const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl();
2580     const FieldDecl *FD = VarFieldMap.lookup(VD);
2581     LValue GlobLVal = CGF.EmitLValueForField(
2582         CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD);
2583     Address GlobAddr = GlobLVal.getAddress(CGF);
2584     llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(
2585         GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs);
2586     llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr);
2587     CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy);
2588     if ((*IPriv)->getType()->isVariablyModifiedType()) {
2589       // Store array size.
2590       ++Idx;
2591       Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2592       llvm::Value *Size = CGF.Builder.CreateIntCast(
2593           CGF.getVLASize(
2594                  CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
2595               .NumElts,
2596           CGF.SizeTy, /*isSigned=*/false);
2597       CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
2598                               Elem);
2599     }
2600   }
2601 
2602   // Call reduce_function(ReduceList, GlobalReduceList)
2603   llvm::Value *GlobalReduceList =
2604       CGF.EmitCastToVoidPtr(ReductionList.getPointer());
2605   Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg);
2606   llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar(
2607       AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc);
2608   CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
2609       CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList});
2610   CGF.FinishFunction();
2611   return Fn;
2612 }
2613 
2614 ///
2615 /// Design of OpenMP reductions on the GPU
2616 ///
2617 /// Consider a typical OpenMP program with one or more reduction
2618 /// clauses:
2619 ///
2620 /// float foo;
2621 /// double bar;
2622 /// #pragma omp target teams distribute parallel for \
2623 ///             reduction(+:foo) reduction(*:bar)
2624 /// for (int i = 0; i < N; i++) {
2625 ///   foo += A[i]; bar *= B[i];
2626 /// }
2627 ///
2628 /// where 'foo' and 'bar' are reduced across all OpenMP threads in
2629 /// all teams.  In our OpenMP implementation on the NVPTX device an
2630 /// OpenMP team is mapped to a CUDA threadblock and OpenMP threads
2631 /// within a team are mapped to CUDA threads within a threadblock.
2632 /// Our goal is to efficiently aggregate values across all OpenMP
2633 /// threads such that:
2634 ///
2635 ///   - the compiler and runtime are logically concise, and
2636 ///   - the reduction is performed efficiently in a hierarchical
2637 ///     manner as follows: within OpenMP threads in the same warp,
2638 ///     across warps in a threadblock, and finally across teams on
2639 ///     the NVPTX device.
2640 ///
2641 /// Introduction to Decoupling
2642 ///
2643 /// We would like to decouple the compiler and the runtime so that the
2644 /// latter is ignorant of the reduction variables (number, data types)
2645 /// and the reduction operators.  This allows a simpler interface
2646 /// and implementation while still attaining good performance.
2647 ///
2648 /// Pseudocode for the aforementioned OpenMP program generated by the
2649 /// compiler is as follows:
2650 ///
2651 /// 1. Create private copies of reduction variables on each OpenMP
2652 ///    thread: 'foo_private', 'bar_private'
2653 /// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned
2654 ///    to it and writes the result in 'foo_private' and 'bar_private'
2655 ///    respectively.
2656 /// 3. Call the OpenMP runtime on the GPU to reduce within a team
2657 ///    and store the result on the team master:
2658 ///
2659 ///     __kmpc_nvptx_parallel_reduce_nowait_v2(...,
2660 ///        reduceData, shuffleReduceFn, interWarpCpyFn)
2661 ///
2662 ///     where:
2663 ///       struct ReduceData {
2664 ///         double *foo;
2665 ///         double *bar;
2666 ///       } reduceData
2667 ///       reduceData.foo = &foo_private
2668 ///       reduceData.bar = &bar_private
2669 ///
2670 ///     'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two
2671 ///     auxiliary functions generated by the compiler that operate on
2672 ///     variables of type 'ReduceData'.  They aid the runtime perform
2673 ///     algorithmic steps in a data agnostic manner.
2674 ///
2675 ///     'shuffleReduceFn' is a pointer to a function that reduces data
2676 ///     of type 'ReduceData' across two OpenMP threads (lanes) in the
2677 ///     same warp.  It takes the following arguments as input:
2678 ///
2679 ///     a. variable of type 'ReduceData' on the calling lane,
2680 ///     b. its lane_id,
2681 ///     c. an offset relative to the current lane_id to generate a
2682 ///        remote_lane_id.  The remote lane contains the second
2683 ///        variable of type 'ReduceData' that is to be reduced.
2684 ///     d. an algorithm version parameter determining which reduction
2685 ///        algorithm to use.
2686 ///
2687 ///     'shuffleReduceFn' retrieves data from the remote lane using
2688 ///     efficient GPU shuffle intrinsics and reduces, using the
2689 ///     algorithm specified by the 4th parameter, the two operands
2690 ///     element-wise.  The result is written to the first operand.
2691 ///
2692 ///     Different reduction algorithms are implemented in different
2693 ///     runtime functions, all calling 'shuffleReduceFn' to perform
2694 ///     the essential reduction step.  Therefore, based on the 4th
2695 ///     parameter, this function behaves slightly differently to
2696 ///     cooperate with the runtime to ensure correctness under
2697 ///     different circumstances.
2698 ///
2699 ///     'InterWarpCpyFn' is a pointer to a function that transfers
2700 ///     reduced variables across warps.  It tunnels, through CUDA
2701 ///     shared memory, the thread-private data of type 'ReduceData'
2702 ///     from lane 0 of each warp to a lane in the first warp.
2703 /// 4. Call the OpenMP runtime on the GPU to reduce across teams.
2704 ///    The last team writes the global reduced value to memory.
2705 ///
2706 ///     ret = __kmpc_nvptx_teams_reduce_nowait(...,
2707 ///             reduceData, shuffleReduceFn, interWarpCpyFn,
2708 ///             scratchpadCopyFn, loadAndReduceFn)
2709 ///
2710 ///     'scratchpadCopyFn' is a helper that stores reduced
2711 ///     data from the team master to a scratchpad array in
2712 ///     global memory.
2713 ///
2714 ///     'loadAndReduceFn' is a helper that loads data from
2715 ///     the scratchpad array and reduces it with the input
2716 ///     operand.
2717 ///
2718 ///     These compiler generated functions hide address
2719 ///     calculation and alignment information from the runtime.
2720 /// 5. if ret == 1:
2721 ///     The team master of the last team stores the reduced
2722 ///     result to the globals in memory.
2723 ///     foo += reduceData.foo; bar *= reduceData.bar
2724 ///
2725 ///
2726 /// Warp Reduction Algorithms
2727 ///
2728 /// On the warp level, we have three algorithms implemented in the
2729 /// OpenMP runtime depending on the number of active lanes:
2730 ///
2731 /// Full Warp Reduction
2732 ///
2733 /// The reduce algorithm within a warp where all lanes are active
2734 /// is implemented in the runtime as follows:
2735 ///
2736 /// full_warp_reduce(void *reduce_data,
2737 ///                  kmp_ShuffleReductFctPtr ShuffleReduceFn) {
2738 ///   for (int offset = WARPSIZE/2; offset > 0; offset /= 2)
2739 ///     ShuffleReduceFn(reduce_data, 0, offset, 0);
2740 /// }
2741 ///
2742 /// The algorithm completes in log(2, WARPSIZE) steps.
2743 ///
2744 /// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is
2745 /// not used therefore we save instructions by not retrieving lane_id
2746 /// from the corresponding special registers.  The 4th parameter, which
2747 /// represents the version of the algorithm being used, is set to 0 to
2748 /// signify full warp reduction.
2749 ///
2750 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
2751 ///
2752 /// #reduce_elem refers to an element in the local lane's data structure
2753 /// #remote_elem is retrieved from a remote lane
2754 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
2755 /// reduce_elem = reduce_elem REDUCE_OP remote_elem;
2756 ///
2757 /// Contiguous Partial Warp Reduction
2758 ///
2759 /// This reduce algorithm is used within a warp where only the first
2760 /// 'n' (n <= WARPSIZE) lanes are active.  It is typically used when the
2761 /// number of OpenMP threads in a parallel region is not a multiple of
2762 /// WARPSIZE.  The algorithm is implemented in the runtime as follows:
2763 ///
2764 /// void
2765 /// contiguous_partial_reduce(void *reduce_data,
2766 ///                           kmp_ShuffleReductFctPtr ShuffleReduceFn,
2767 ///                           int size, int lane_id) {
2768 ///   int curr_size;
2769 ///   int offset;
2770 ///   curr_size = size;
2771 ///   mask = curr_size/2;
2772 ///   while (offset>0) {
2773 ///     ShuffleReduceFn(reduce_data, lane_id, offset, 1);
2774 ///     curr_size = (curr_size+1)/2;
2775 ///     offset = curr_size/2;
2776 ///   }
2777 /// }
2778 ///
2779 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
2780 ///
2781 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
2782 /// if (lane_id < offset)
2783 ///     reduce_elem = reduce_elem REDUCE_OP remote_elem
2784 /// else
2785 ///     reduce_elem = remote_elem
2786 ///
2787 /// This algorithm assumes that the data to be reduced are located in a
2788 /// contiguous subset of lanes starting from the first.  When there is
2789 /// an odd number of active lanes, the data in the last lane is not
2790 /// aggregated with any other lane's dat but is instead copied over.
2791 ///
2792 /// Dispersed Partial Warp Reduction
2793 ///
2794 /// This algorithm is used within a warp when any discontiguous subset of
2795 /// lanes are active.  It is used to implement the reduction operation
2796 /// across lanes in an OpenMP simd region or in a nested parallel region.
2797 ///
2798 /// void
2799 /// dispersed_partial_reduce(void *reduce_data,
2800 ///                          kmp_ShuffleReductFctPtr ShuffleReduceFn) {
2801 ///   int size, remote_id;
2802 ///   int logical_lane_id = number_of_active_lanes_before_me() * 2;
2803 ///   do {
2804 ///       remote_id = next_active_lane_id_right_after_me();
2805 ///       # the above function returns 0 of no active lane
2806 ///       # is present right after the current lane.
2807 ///       size = number_of_active_lanes_in_this_warp();
2808 ///       logical_lane_id /= 2;
2809 ///       ShuffleReduceFn(reduce_data, logical_lane_id,
2810 ///                       remote_id-1-threadIdx.x, 2);
2811 ///   } while (logical_lane_id % 2 == 0 && size > 1);
2812 /// }
2813 ///
2814 /// There is no assumption made about the initial state of the reduction.
2815 /// Any number of lanes (>=1) could be active at any position.  The reduction
2816 /// result is returned in the first active lane.
2817 ///
2818 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows:
2819 ///
2820 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE);
2821 /// if (lane_id % 2 == 0 && offset > 0)
2822 ///     reduce_elem = reduce_elem REDUCE_OP remote_elem
2823 /// else
2824 ///     reduce_elem = remote_elem
2825 ///
2826 ///
2827 /// Intra-Team Reduction
2828 ///
2829 /// This function, as implemented in the runtime call
2830 /// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP
2831 /// threads in a team.  It first reduces within a warp using the
2832 /// aforementioned algorithms.  We then proceed to gather all such
2833 /// reduced values at the first warp.
2834 ///
2835 /// The runtime makes use of the function 'InterWarpCpyFn', which copies
2836 /// data from each of the "warp master" (zeroth lane of each warp, where
2837 /// warp-reduced data is held) to the zeroth warp.  This step reduces (in
2838 /// a mathematical sense) the problem of reduction across warp masters in
2839 /// a block to the problem of warp reduction.
2840 ///
2841 ///
2842 /// Inter-Team Reduction
2843 ///
2844 /// Once a team has reduced its data to a single value, it is stored in
2845 /// a global scratchpad array.  Since each team has a distinct slot, this
2846 /// can be done without locking.
2847 ///
2848 /// The last team to write to the scratchpad array proceeds to reduce the
2849 /// scratchpad array.  One or more workers in the last team use the helper
2850 /// 'loadAndReduceDataFn' to load and reduce values from the array, i.e.,
2851 /// the k'th worker reduces every k'th element.
2852 ///
2853 /// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to
2854 /// reduce across workers and compute a globally reduced value.
2855 ///
2856 void CGOpenMPRuntimeGPU::emitReduction(
2857     CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates,
2858     ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs,
2859     ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) {
2860   if (!CGF.HaveInsertPoint())
2861     return;
2862 
2863   bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind);
2864 #ifndef NDEBUG
2865   bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind);
2866 #endif
2867 
2868   if (Options.SimpleReduction) {
2869     assert(!TeamsReduction && !ParallelReduction &&
2870            "Invalid reduction selection in emitReduction.");
2871     CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs,
2872                                    ReductionOps, Options);
2873     return;
2874   }
2875 
2876   assert((TeamsReduction || ParallelReduction) &&
2877          "Invalid reduction selection in emitReduction.");
2878 
2879   // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList),
2880   // RedList, shuffle_reduce_func, interwarp_copy_func);
2881   // or
2882   // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>);
2883   llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc);
2884   llvm::Value *ThreadId = getThreadID(CGF, Loc);
2885 
2886   llvm::Value *Res;
2887   ASTContext &C = CGM.getContext();
2888   // 1. Build a list of reduction variables.
2889   // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
2890   auto Size = RHSExprs.size();
2891   for (const Expr *E : Privates) {
2892     if (E->getType()->isVariablyModifiedType())
2893       // Reserve place for array size.
2894       ++Size;
2895   }
2896   llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size);
2897   QualType ReductionArrayTy =
2898       C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal,
2899                              /*IndexTypeQuals=*/0);
2900   Address ReductionList =
2901       CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list");
2902   auto IPriv = Privates.begin();
2903   unsigned Idx = 0;
2904   for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) {
2905     Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2906     CGF.Builder.CreateStore(
2907         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2908             CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy),
2909         Elem);
2910     if ((*IPriv)->getType()->isVariablyModifiedType()) {
2911       // Store array size.
2912       ++Idx;
2913       Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx);
2914       llvm::Value *Size = CGF.Builder.CreateIntCast(
2915           CGF.getVLASize(
2916                  CGF.getContext().getAsVariableArrayType((*IPriv)->getType()))
2917               .NumElts,
2918           CGF.SizeTy, /*isSigned=*/false);
2919       CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy),
2920                               Elem);
2921     }
2922   }
2923 
2924   llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
2925       ReductionList.getPointer(), CGF.VoidPtrTy);
2926   llvm::Function *ReductionFn =
2927       emitReductionFunction(Loc, CGF.ConvertTypeForMem(ReductionArrayTy),
2928                             Privates, LHSExprs, RHSExprs, ReductionOps);
2929   llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy);
2930   llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction(
2931       CGM, Privates, ReductionArrayTy, ReductionFn, Loc);
2932   llvm::Value *InterWarpCopyFn =
2933       emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc);
2934 
2935   if (ParallelReduction) {
2936     llvm::Value *Args[] = {RTLoc,
2937                            ThreadId,
2938                            CGF.Builder.getInt32(RHSExprs.size()),
2939                            ReductionArrayTySize,
2940                            RL,
2941                            ShuffleAndReduceFn,
2942                            InterWarpCopyFn};
2943 
2944     Res = CGF.EmitRuntimeCall(
2945         OMPBuilder.getOrCreateRuntimeFunction(
2946             CGM.getModule(), OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2),
2947         Args);
2948   } else {
2949     assert(TeamsReduction && "expected teams reduction.");
2950     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap;
2951     llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size());
2952     int Cnt = 0;
2953     for (const Expr *DRE : Privates) {
2954       PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl();
2955       ++Cnt;
2956     }
2957     const RecordDecl *TeamReductionRec = ::buildRecordForGlobalizedVars(
2958         CGM.getContext(), PrivatesReductions, std::nullopt, VarFieldMap,
2959         C.getLangOpts().OpenMPCUDAReductionBufNum);
2960     TeamsReductions.push_back(TeamReductionRec);
2961     if (!KernelTeamsReductionPtr) {
2962       KernelTeamsReductionPtr = new llvm::GlobalVariable(
2963           CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/true,
2964           llvm::GlobalValue::InternalLinkage, nullptr,
2965           "_openmp_teams_reductions_buffer_$_$ptr");
2966     }
2967     llvm::Value *GlobalBufferPtr = CGF.EmitLoadOfScalar(
2968         Address(KernelTeamsReductionPtr, CGF.VoidPtrTy, CGM.getPointerAlign()),
2969         /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc);
2970     llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction(
2971         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap);
2972     llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction(
2973         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap,
2974         ReductionFn);
2975     llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction(
2976         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap);
2977     llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction(
2978         CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap,
2979         ReductionFn);
2980 
2981     llvm::Value *Args[] = {
2982         RTLoc,
2983         ThreadId,
2984         GlobalBufferPtr,
2985         CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum),
2986         RL,
2987         ShuffleAndReduceFn,
2988         InterWarpCopyFn,
2989         GlobalToBufferCpyFn,
2990         GlobalToBufferRedFn,
2991         BufferToGlobalCpyFn,
2992         BufferToGlobalRedFn};
2993 
2994     Res = CGF.EmitRuntimeCall(
2995         OMPBuilder.getOrCreateRuntimeFunction(
2996             CGM.getModule(), OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2),
2997         Args);
2998   }
2999 
3000   // 5. Build if (res == 1)
3001   llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done");
3002   llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then");
3003   llvm::Value *Cond = CGF.Builder.CreateICmpEQ(
3004       Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1));
3005   CGF.Builder.CreateCondBr(Cond, ThenBB, ExitBB);
3006 
3007   // 6. Build then branch: where we have reduced values in the master
3008   //    thread in each team.
3009   //    __kmpc_end_reduce{_nowait}(<gtid>);
3010   //    break;
3011   CGF.EmitBlock(ThenBB);
3012 
3013   // Add emission of __kmpc_end_reduce{_nowait}(<gtid>);
3014   auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps,
3015                     this](CodeGenFunction &CGF, PrePostActionTy &Action) {
3016     auto IPriv = Privates.begin();
3017     auto ILHS = LHSExprs.begin();
3018     auto IRHS = RHSExprs.begin();
3019     for (const Expr *E : ReductionOps) {
3020       emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS),
3021                                   cast<DeclRefExpr>(*IRHS));
3022       ++IPriv;
3023       ++ILHS;
3024       ++IRHS;
3025     }
3026   };
3027   llvm::Value *EndArgs[] = {ThreadId};
3028   RegionCodeGenTy RCG(CodeGen);
3029   NVPTXActionTy Action(
3030       nullptr, std::nullopt,
3031       OMPBuilder.getOrCreateRuntimeFunction(
3032           CGM.getModule(), OMPRTL___kmpc_nvptx_end_reduce_nowait),
3033       EndArgs);
3034   RCG.setAction(Action);
3035   RCG(CGF);
3036   // There is no need to emit line number for unconditional branch.
3037   (void)ApplyDebugLocation::CreateEmpty(CGF);
3038   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
3039 }
3040 
3041 const VarDecl *
3042 CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD,
3043                                        const VarDecl *NativeParam) const {
3044   if (!NativeParam->getType()->isReferenceType())
3045     return NativeParam;
3046   QualType ArgType = NativeParam->getType();
3047   QualifierCollector QC;
3048   const Type *NonQualTy = QC.strip(ArgType);
3049   QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
3050   if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) {
3051     if (Attr->getCaptureKind() == OMPC_map) {
3052       PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy,
3053                                                         LangAS::opencl_global);
3054     }
3055   }
3056   ArgType = CGM.getContext().getPointerType(PointeeTy);
3057   QC.addRestrict();
3058   enum { NVPTX_local_addr = 5 };
3059   QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr));
3060   ArgType = QC.apply(CGM.getContext(), ArgType);
3061   if (isa<ImplicitParamDecl>(NativeParam))
3062     return ImplicitParamDecl::Create(
3063         CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(),
3064         NativeParam->getIdentifier(), ArgType, ImplicitParamDecl::Other);
3065   return ParmVarDecl::Create(
3066       CGM.getContext(),
3067       const_cast<DeclContext *>(NativeParam->getDeclContext()),
3068       NativeParam->getBeginLoc(), NativeParam->getLocation(),
3069       NativeParam->getIdentifier(), ArgType,
3070       /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr);
3071 }
3072 
3073 Address
3074 CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF,
3075                                           const VarDecl *NativeParam,
3076                                           const VarDecl *TargetParam) const {
3077   assert(NativeParam != TargetParam &&
3078          NativeParam->getType()->isReferenceType() &&
3079          "Native arg must not be the same as target arg.");
3080   Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam);
3081   QualType NativeParamType = NativeParam->getType();
3082   QualifierCollector QC;
3083   const Type *NonQualTy = QC.strip(NativeParamType);
3084   QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType();
3085   unsigned NativePointeeAddrSpace =
3086       CGF.getTypes().getTargetAddressSpace(NativePointeeTy);
3087   QualType TargetTy = TargetParam->getType();
3088   llvm::Value *TargetAddr = CGF.EmitLoadOfScalar(
3089       LocalAddr, /*Volatile=*/false, TargetTy, SourceLocation());
3090   // First cast to generic.
3091   TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3092       TargetAddr, llvm::PointerType::getWithSamePointeeType(
3093           cast<llvm::PointerType>(TargetAddr->getType()), /*AddrSpace=*/0));
3094   // Cast from generic to native address space.
3095   TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3096       TargetAddr, llvm::PointerType::getWithSamePointeeType(
3097           cast<llvm::PointerType>(TargetAddr->getType()),
3098                                   NativePointeeAddrSpace));
3099   Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType);
3100   CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false,
3101                         NativeParamType);
3102   return NativeParamAddr;
3103 }
3104 
3105 void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall(
3106     CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn,
3107     ArrayRef<llvm::Value *> Args) const {
3108   SmallVector<llvm::Value *, 4> TargetArgs;
3109   TargetArgs.reserve(Args.size());
3110   auto *FnType = OutlinedFn.getFunctionType();
3111   for (unsigned I = 0, E = Args.size(); I < E; ++I) {
3112     if (FnType->isVarArg() && FnType->getNumParams() <= I) {
3113       TargetArgs.append(std::next(Args.begin(), I), Args.end());
3114       break;
3115     }
3116     llvm::Type *TargetType = FnType->getParamType(I);
3117     llvm::Value *NativeArg = Args[I];
3118     if (!TargetType->isPointerTy()) {
3119       TargetArgs.emplace_back(NativeArg);
3120       continue;
3121     }
3122     llvm::Value *TargetArg = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3123         NativeArg, llvm::PointerType::getWithSamePointeeType(
3124             cast<llvm::PointerType>(NativeArg->getType()), /*AddrSpace*/ 0));
3125     TargetArgs.emplace_back(
3126         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType));
3127   }
3128   CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs);
3129 }
3130 
3131 /// Emit function which wraps the outline parallel region
3132 /// and controls the arguments which are passed to this function.
3133 /// The wrapper ensures that the outlined function is called
3134 /// with the correct arguments when data is shared.
3135 llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper(
3136     llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) {
3137   ASTContext &Ctx = CGM.getContext();
3138   const auto &CS = *D.getCapturedStmt(OMPD_parallel);
3139 
3140   // Create a function that takes as argument the source thread.
3141   FunctionArgList WrapperArgs;
3142   QualType Int16QTy =
3143       Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false);
3144   QualType Int32QTy =
3145       Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false);
3146   ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
3147                                      /*Id=*/nullptr, Int16QTy,
3148                                      ImplicitParamDecl::Other);
3149   ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(),
3150                                /*Id=*/nullptr, Int32QTy,
3151                                ImplicitParamDecl::Other);
3152   WrapperArgs.emplace_back(&ParallelLevelArg);
3153   WrapperArgs.emplace_back(&WrapperArg);
3154 
3155   const CGFunctionInfo &CGFI =
3156       CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs);
3157 
3158   auto *Fn = llvm::Function::Create(
3159       CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage,
3160       Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule());
3161 
3162   // Ensure we do not inline the function. This is trivially true for the ones
3163   // passed to __kmpc_fork_call but the ones calles in serialized regions
3164   // could be inlined. This is not a perfect but it is closer to the invariant
3165   // we want, namely, every data environment starts with a new function.
3166   // TODO: We should pass the if condition to the runtime function and do the
3167   //       handling there. Much cleaner code.
3168   Fn->addFnAttr(llvm::Attribute::NoInline);
3169 
3170   CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI);
3171   Fn->setLinkage(llvm::GlobalValue::InternalLinkage);
3172   Fn->setDoesNotRecurse();
3173 
3174   CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
3175   CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs,
3176                     D.getBeginLoc(), D.getBeginLoc());
3177 
3178   const auto *RD = CS.getCapturedRecordDecl();
3179   auto CurField = RD->field_begin();
3180 
3181   Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty,
3182                                                       /*Name=*/".zero.addr");
3183   CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr);
3184   // Get the array of arguments.
3185   SmallVector<llvm::Value *, 8> Args;
3186 
3187   Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer());
3188   Args.emplace_back(ZeroAddr.getPointer());
3189 
3190   CGBuilderTy &Bld = CGF.Builder;
3191   auto CI = CS.capture_begin();
3192 
3193   // Use global memory for data sharing.
3194   // Handle passing of global args to workers.
3195   Address GlobalArgs =
3196       CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args");
3197   llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer();
3198   llvm::Value *DataSharingArgs[] = {GlobalArgsPtr};
3199   CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
3200                           CGM.getModule(), OMPRTL___kmpc_get_shared_variables),
3201                       DataSharingArgs);
3202 
3203   // Retrieve the shared variables from the list of references returned
3204   // by the runtime. Pass the variables to the outlined function.
3205   Address SharedArgListAddress = Address::invalid();
3206   if (CS.capture_size() > 0 ||
3207       isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
3208     SharedArgListAddress = CGF.EmitLoadOfPointer(
3209         GlobalArgs, CGF.getContext()
3210                         .getPointerType(CGF.getContext().VoidPtrTy)
3211                         .castAs<PointerType>());
3212   }
3213   unsigned Idx = 0;
3214   if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) {
3215     Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
3216     Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
3217         Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy);
3218     llvm::Value *LB = CGF.EmitLoadOfScalar(
3219         TypedAddress,
3220         /*Volatile=*/false,
3221         CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
3222         cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc());
3223     Args.emplace_back(LB);
3224     ++Idx;
3225     Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx);
3226     TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
3227         Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy);
3228     llvm::Value *UB = CGF.EmitLoadOfScalar(
3229         TypedAddress,
3230         /*Volatile=*/false,
3231         CGF.getContext().getPointerType(CGF.getContext().getSizeType()),
3232         cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc());
3233     Args.emplace_back(UB);
3234     ++Idx;
3235   }
3236   if (CS.capture_size() > 0) {
3237     ASTContext &CGFContext = CGF.getContext();
3238     for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) {
3239       QualType ElemTy = CurField->getType();
3240       Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx);
3241       Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(
3242           Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)),
3243           CGF.ConvertTypeForMem(ElemTy));
3244       llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress,
3245                                               /*Volatile=*/false,
3246                                               CGFContext.getPointerType(ElemTy),
3247                                               CI->getLocation());
3248       if (CI->capturesVariableByCopy() &&
3249           !CI->getCapturedVar()->getType()->isAnyPointerType()) {
3250         Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(),
3251                               CI->getLocation());
3252       }
3253       Args.emplace_back(Arg);
3254     }
3255   }
3256 
3257   emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args);
3258   CGF.FinishFunction();
3259   return Fn;
3260 }
3261 
3262 void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF,
3263                                               const Decl *D) {
3264   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic)
3265     return;
3266 
3267   assert(D && "Expected function or captured|block decl.");
3268   assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 &&
3269          "Function is registered already.");
3270   assert((!TeamAndReductions.first || TeamAndReductions.first == D) &&
3271          "Team is set but not processed.");
3272   const Stmt *Body = nullptr;
3273   bool NeedToDelayGlobalization = false;
3274   if (const auto *FD = dyn_cast<FunctionDecl>(D)) {
3275     Body = FD->getBody();
3276   } else if (const auto *BD = dyn_cast<BlockDecl>(D)) {
3277     Body = BD->getBody();
3278   } else if (const auto *CD = dyn_cast<CapturedDecl>(D)) {
3279     Body = CD->getBody();
3280     NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP;
3281     if (NeedToDelayGlobalization &&
3282         getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD)
3283       return;
3284   }
3285   if (!Body)
3286     return;
3287   CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second);
3288   VarChecker.Visit(Body);
3289   const RecordDecl *GlobalizedVarsRecord =
3290       VarChecker.getGlobalizedRecord(IsInTTDRegion);
3291   TeamAndReductions.first = nullptr;
3292   TeamAndReductions.second.clear();
3293   ArrayRef<const ValueDecl *> EscapedVariableLengthDecls =
3294       VarChecker.getEscapedVariableLengthDecls();
3295   if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty())
3296     return;
3297   auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first;
3298   I->getSecond().MappedParams =
3299       std::make_unique<CodeGenFunction::OMPMapVars>();
3300   I->getSecond().EscapedParameters.insert(
3301       VarChecker.getEscapedParameters().begin(),
3302       VarChecker.getEscapedParameters().end());
3303   I->getSecond().EscapedVariableLengthDecls.append(
3304       EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end());
3305   DeclToAddrMapTy &Data = I->getSecond().LocalVarData;
3306   for (const ValueDecl *VD : VarChecker.getEscapedDecls()) {
3307     assert(VD->isCanonicalDecl() && "Expected canonical declaration");
3308     Data.insert(std::make_pair(VD, MappedVarData()));
3309   }
3310   if (!NeedToDelayGlobalization) {
3311     emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true);
3312     struct GlobalizationScope final : EHScopeStack::Cleanup {
3313       GlobalizationScope() = default;
3314 
3315       void Emit(CodeGenFunction &CGF, Flags flags) override {
3316         static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime())
3317             .emitGenericVarsEpilog(CGF, /*WithSPMDCheck=*/true);
3318       }
3319     };
3320     CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup);
3321   }
3322 }
3323 
3324 Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF,
3325                                                         const VarDecl *VD) {
3326   if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) {
3327     const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
3328     auto AS = LangAS::Default;
3329     switch (A->getAllocatorType()) {
3330       // Use the default allocator here as by default local vars are
3331       // threadlocal.
3332     case OMPAllocateDeclAttr::OMPNullMemAlloc:
3333     case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
3334     case OMPAllocateDeclAttr::OMPThreadMemAlloc:
3335     case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
3336     case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
3337       // Follow the user decision - use default allocation.
3338       return Address::invalid();
3339     case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
3340       // TODO: implement aupport for user-defined allocators.
3341       return Address::invalid();
3342     case OMPAllocateDeclAttr::OMPConstMemAlloc:
3343       AS = LangAS::cuda_constant;
3344       break;
3345     case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
3346       AS = LangAS::cuda_shared;
3347       break;
3348     case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
3349     case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
3350       break;
3351     }
3352     llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
3353     auto *GV = new llvm::GlobalVariable(
3354         CGM.getModule(), VarTy, /*isConstant=*/false,
3355         llvm::GlobalValue::InternalLinkage, llvm::Constant::getNullValue(VarTy),
3356         VD->getName(),
3357         /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
3358         CGM.getContext().getTargetAddressSpace(AS));
3359     CharUnits Align = CGM.getContext().getDeclAlign(VD);
3360     GV->setAlignment(Align.getAsAlign());
3361     return Address(
3362         CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
3363             GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace(
3364                     VD->getType().getAddressSpace()))),
3365         VarTy, Align);
3366   }
3367 
3368   if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic)
3369     return Address::invalid();
3370 
3371   VD = VD->getCanonicalDecl();
3372   auto I = FunctionGlobalizedDecls.find(CGF.CurFn);
3373   if (I == FunctionGlobalizedDecls.end())
3374     return Address::invalid();
3375   auto VDI = I->getSecond().LocalVarData.find(VD);
3376   if (VDI != I->getSecond().LocalVarData.end())
3377     return VDI->second.PrivateAddr;
3378   if (VD->hasAttrs()) {
3379     for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()),
3380          E(VD->attr_end());
3381          IT != E; ++IT) {
3382       auto VDI = I->getSecond().LocalVarData.find(
3383           cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl())
3384               ->getCanonicalDecl());
3385       if (VDI != I->getSecond().LocalVarData.end())
3386         return VDI->second.PrivateAddr;
3387     }
3388   }
3389 
3390   return Address::invalid();
3391 }
3392 
3393 void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) {
3394   FunctionGlobalizedDecls.erase(CGF.CurFn);
3395   CGOpenMPRuntime::functionFinished(CGF);
3396 }
3397 
3398 void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk(
3399     CodeGenFunction &CGF, const OMPLoopDirective &S,
3400     OpenMPDistScheduleClauseKind &ScheduleKind,
3401     llvm::Value *&Chunk) const {
3402   auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime());
3403   if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) {
3404     ScheduleKind = OMPC_DIST_SCHEDULE_static;
3405     Chunk = CGF.EmitScalarConversion(
3406         RT.getGPUNumThreads(CGF),
3407         CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
3408         S.getIterationVariable()->getType(), S.getBeginLoc());
3409     return;
3410   }
3411   CGOpenMPRuntime::getDefaultDistScheduleAndChunk(
3412       CGF, S, ScheduleKind, Chunk);
3413 }
3414 
3415 void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk(
3416     CodeGenFunction &CGF, const OMPLoopDirective &S,
3417     OpenMPScheduleClauseKind &ScheduleKind,
3418     const Expr *&ChunkExpr) const {
3419   ScheduleKind = OMPC_SCHEDULE_static;
3420   // Chunk size is 1 in this case.
3421   llvm::APInt ChunkSize(32, 1);
3422   ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize,
3423       CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
3424       SourceLocation());
3425 }
3426 
3427 void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas(
3428     CodeGenFunction &CGF, const OMPExecutableDirective &D) const {
3429   assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&
3430          " Expected target-based directive.");
3431   const CapturedStmt *CS = D.getCapturedStmt(OMPD_target);
3432   for (const CapturedStmt::Capture &C : CS->captures()) {
3433     // Capture variables captured by reference in lambdas for target-based
3434     // directives.
3435     if (!C.capturesVariable())
3436       continue;
3437     const VarDecl *VD = C.getCapturedVar();
3438     const auto *RD = VD->getType()
3439                          .getCanonicalType()
3440                          .getNonReferenceType()
3441                          ->getAsCXXRecordDecl();
3442     if (!RD || !RD->isLambda())
3443       continue;
3444     Address VDAddr = CGF.GetAddrOfLocalVar(VD);
3445     LValue VDLVal;
3446     if (VD->getType().getCanonicalType()->isReferenceType())
3447       VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType());
3448     else
3449       VDLVal = CGF.MakeAddrLValue(
3450           VDAddr, VD->getType().getCanonicalType().getNonReferenceType());
3451     llvm::DenseMap<const ValueDecl *, FieldDecl *> Captures;
3452     FieldDecl *ThisCapture = nullptr;
3453     RD->getCaptureFields(Captures, ThisCapture);
3454     if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) {
3455       LValue ThisLVal =
3456           CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture);
3457       llvm::Value *CXXThis = CGF.LoadCXXThis();
3458       CGF.EmitStoreOfScalar(CXXThis, ThisLVal);
3459     }
3460     for (const LambdaCapture &LC : RD->captures()) {
3461       if (LC.getCaptureKind() != LCK_ByRef)
3462         continue;
3463       const ValueDecl *VD = LC.getCapturedVar();
3464       // FIXME: For now VD is always a VarDecl because OpenMP does not support
3465       //  capturing structured bindings in lambdas yet.
3466       if (!CS->capturesVariable(cast<VarDecl>(VD)))
3467         continue;
3468       auto It = Captures.find(VD);
3469       assert(It != Captures.end() && "Found lambda capture without field.");
3470       LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second);
3471       Address VDAddr = CGF.GetAddrOfLocalVar(cast<VarDecl>(VD));
3472       if (VD->getType().getCanonicalType()->isReferenceType())
3473         VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr,
3474                                                VD->getType().getCanonicalType())
3475                      .getAddress(CGF);
3476       CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal);
3477     }
3478   }
3479 }
3480 
3481 bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD,
3482                                                             LangAS &AS) {
3483   if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())
3484     return false;
3485   const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
3486   switch(A->getAllocatorType()) {
3487   case OMPAllocateDeclAttr::OMPNullMemAlloc:
3488   case OMPAllocateDeclAttr::OMPDefaultMemAlloc:
3489   // Not supported, fallback to the default mem space.
3490   case OMPAllocateDeclAttr::OMPThreadMemAlloc:
3491   case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
3492   case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
3493   case OMPAllocateDeclAttr::OMPHighBWMemAlloc:
3494   case OMPAllocateDeclAttr::OMPLowLatMemAlloc:
3495     AS = LangAS::Default;
3496     return true;
3497   case OMPAllocateDeclAttr::OMPConstMemAlloc:
3498     AS = LangAS::cuda_constant;
3499     return true;
3500   case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
3501     AS = LangAS::cuda_shared;
3502     return true;
3503   case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
3504     llvm_unreachable("Expected predefined allocator for the variables with the "
3505                      "static storage.");
3506   }
3507   return false;
3508 }
3509 
3510 // Get current CudaArch and ignore any unknown values
3511 static CudaArch getCudaArch(CodeGenModule &CGM) {
3512   if (!CGM.getTarget().hasFeature("ptx"))
3513     return CudaArch::UNKNOWN;
3514   for (const auto &Feature : CGM.getTarget().getTargetOpts().FeatureMap) {
3515     if (Feature.getValue()) {
3516       CudaArch Arch = StringToCudaArch(Feature.getKey());
3517       if (Arch != CudaArch::UNKNOWN)
3518         return Arch;
3519     }
3520   }
3521   return CudaArch::UNKNOWN;
3522 }
3523 
3524 /// Check to see if target architecture supports unified addressing which is
3525 /// a restriction for OpenMP requires clause "unified_shared_memory".
3526 void CGOpenMPRuntimeGPU::processRequiresDirective(
3527     const OMPRequiresDecl *D) {
3528   for (const OMPClause *Clause : D->clauselists()) {
3529     if (Clause->getClauseKind() == OMPC_unified_shared_memory) {
3530       CudaArch Arch = getCudaArch(CGM);
3531       switch (Arch) {
3532       case CudaArch::SM_20:
3533       case CudaArch::SM_21:
3534       case CudaArch::SM_30:
3535       case CudaArch::SM_32:
3536       case CudaArch::SM_35:
3537       case CudaArch::SM_37:
3538       case CudaArch::SM_50:
3539       case CudaArch::SM_52:
3540       case CudaArch::SM_53: {
3541         SmallString<256> Buffer;
3542         llvm::raw_svector_ostream Out(Buffer);
3543         Out << "Target architecture " << CudaArchToString(Arch)
3544             << " does not support unified addressing";
3545         CGM.Error(Clause->getBeginLoc(), Out.str());
3546         return;
3547       }
3548       case CudaArch::SM_60:
3549       case CudaArch::SM_61:
3550       case CudaArch::SM_62:
3551       case CudaArch::SM_70:
3552       case CudaArch::SM_72:
3553       case CudaArch::SM_75:
3554       case CudaArch::SM_80:
3555       case CudaArch::SM_86:
3556       case CudaArch::SM_87:
3557       case CudaArch::SM_89:
3558       case CudaArch::SM_90:
3559       case CudaArch::GFX600:
3560       case CudaArch::GFX601:
3561       case CudaArch::GFX602:
3562       case CudaArch::GFX700:
3563       case CudaArch::GFX701:
3564       case CudaArch::GFX702:
3565       case CudaArch::GFX703:
3566       case CudaArch::GFX704:
3567       case CudaArch::GFX705:
3568       case CudaArch::GFX801:
3569       case CudaArch::GFX802:
3570       case CudaArch::GFX803:
3571       case CudaArch::GFX805:
3572       case CudaArch::GFX810:
3573       case CudaArch::GFX900:
3574       case CudaArch::GFX902:
3575       case CudaArch::GFX904:
3576       case CudaArch::GFX906:
3577       case CudaArch::GFX908:
3578       case CudaArch::GFX909:
3579       case CudaArch::GFX90a:
3580       case CudaArch::GFX90c:
3581       case CudaArch::GFX940:
3582       case CudaArch::GFX1010:
3583       case CudaArch::GFX1011:
3584       case CudaArch::GFX1012:
3585       case CudaArch::GFX1013:
3586       case CudaArch::GFX1030:
3587       case CudaArch::GFX1031:
3588       case CudaArch::GFX1032:
3589       case CudaArch::GFX1033:
3590       case CudaArch::GFX1034:
3591       case CudaArch::GFX1035:
3592       case CudaArch::GFX1036:
3593       case CudaArch::GFX1100:
3594       case CudaArch::GFX1101:
3595       case CudaArch::GFX1102:
3596       case CudaArch::GFX1103:
3597       case CudaArch::Generic:
3598       case CudaArch::UNUSED:
3599       case CudaArch::UNKNOWN:
3600         break;
3601       case CudaArch::LAST:
3602         llvm_unreachable("Unexpected Cuda arch.");
3603       }
3604     }
3605   }
3606   CGOpenMPRuntime::processRequiresDirective(D);
3607 }
3608 
3609 void CGOpenMPRuntimeGPU::clear() {
3610 
3611   if (!TeamsReductions.empty()) {
3612     ASTContext &C = CGM.getContext();
3613     RecordDecl *StaticRD = C.buildImplicitRecord(
3614         "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::TTK_Union);
3615     StaticRD->startDefinition();
3616     for (const RecordDecl *TeamReductionRec : TeamsReductions) {
3617       QualType RecTy = C.getRecordType(TeamReductionRec);
3618       auto *Field = FieldDecl::Create(
3619           C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy,
3620           C.getTrivialTypeSourceInfo(RecTy, SourceLocation()),
3621           /*BW=*/nullptr, /*Mutable=*/false,
3622           /*InitStyle=*/ICIS_NoInit);
3623       Field->setAccess(AS_public);
3624       StaticRD->addDecl(Field);
3625     }
3626     StaticRD->completeDefinition();
3627     QualType StaticTy = C.getRecordType(StaticRD);
3628     llvm::Type *LLVMReductionsBufferTy =
3629         CGM.getTypes().ConvertTypeForMem(StaticTy);
3630     // FIXME: nvlink does not handle weak linkage correctly (object with the
3631     // different size are reported as erroneous).
3632     // Restore CommonLinkage as soon as nvlink is fixed.
3633     auto *GV = new llvm::GlobalVariable(
3634         CGM.getModule(), LLVMReductionsBufferTy,
3635         /*isConstant=*/false, llvm::GlobalValue::InternalLinkage,
3636         llvm::Constant::getNullValue(LLVMReductionsBufferTy),
3637         "_openmp_teams_reductions_buffer_$_");
3638     KernelTeamsReductionPtr->setInitializer(
3639         llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV,
3640                                                              CGM.VoidPtrTy));
3641   }
3642   CGOpenMPRuntime::clear();
3643 }
3644 
3645 llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) {
3646   CGBuilderTy &Bld = CGF.Builder;
3647   llvm::Module *M = &CGF.CGM.getModule();
3648   const char *LocSize = "__kmpc_get_hardware_num_threads_in_block";
3649   llvm::Function *F = M->getFunction(LocSize);
3650   if (!F) {
3651     F = llvm::Function::Create(
3652         llvm::FunctionType::get(CGF.Int32Ty, std::nullopt, false),
3653         llvm::GlobalVariable::ExternalLinkage, LocSize, &CGF.CGM.getModule());
3654   }
3655   return Bld.CreateCall(F, std::nullopt, "nvptx_num_threads");
3656 }
3657 
3658 llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) {
3659   ArrayRef<llvm::Value *> Args{};
3660   return CGF.EmitRuntimeCall(
3661       OMPBuilder.getOrCreateRuntimeFunction(
3662           CGM.getModule(), OMPRTL___kmpc_get_hardware_thread_id_in_block),
3663       Args);
3664 }
3665 
3666 llvm::Value *CGOpenMPRuntimeGPU::getGPUWarpSize(CodeGenFunction &CGF) {
3667   ArrayRef<llvm::Value *> Args{};
3668   return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
3669                                  CGM.getModule(), OMPRTL___kmpc_get_warp_size),
3670                              Args);
3671 }
3672