1 //===-- OpenMP.cpp -- Open MP directive lowering --------------------------===//
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 // Coding style: https://mlir.llvm.org/getting_started/DeveloperGuide/
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "flang/Lower/OpenMP.h"
14 #include "flang/Common/idioms.h"
15 #include "flang/Lower/Bridge.h"
16 #include "flang/Lower/ConvertExpr.h"
17 #include "flang/Lower/PFTBuilder.h"
18 #include "flang/Lower/StatementContext.h"
19 #include "flang/Optimizer/Builder/BoxValue.h"
20 #include "flang/Optimizer/Builder/FIRBuilder.h"
21 #include "flang/Optimizer/Builder/Todo.h"
22 #include "flang/Parser/parse-tree.h"
23 #include "flang/Semantics/tools.h"
24 #include "mlir/Dialect/OpenMP/OpenMPDialect.h"
25 #include "llvm/Frontend/OpenMP/OMPConstants.h"
26 
27 using namespace mlir;
28 
29 int64_t Fortran::lower::getCollapseValue(
30     const Fortran::parser::OmpClauseList &clauseList) {
31   for (const auto &clause : clauseList.v) {
32     if (const auto &collapseClause =
33             std::get_if<Fortran::parser::OmpClause::Collapse>(&clause.u)) {
34       const auto *expr = Fortran::semantics::GetExpr(collapseClause->v);
35       return Fortran::evaluate::ToInt64(*expr).value();
36     }
37   }
38   return 1;
39 }
40 
41 static const Fortran::parser::Name *
42 getDesignatorNameIfDataRef(const Fortran::parser::Designator &designator) {
43   const auto *dataRef = std::get_if<Fortran::parser::DataRef>(&designator.u);
44   return dataRef ? std::get_if<Fortran::parser::Name>(&dataRef->u) : nullptr;
45 }
46 
47 static Fortran::semantics::Symbol *
48 getOmpObjectSymbol(const Fortran::parser::OmpObject &ompObject) {
49   Fortran::semantics::Symbol *sym = nullptr;
50   std::visit(Fortran::common::visitors{
51                  [&](const Fortran::parser::Designator &designator) {
52                    if (const Fortran::parser::Name *name =
53                            getDesignatorNameIfDataRef(designator)) {
54                      sym = name->symbol;
55                    }
56                  },
57                  [&](const Fortran::parser::Name &name) { sym = name.symbol; }},
58              ompObject.u);
59   return sym;
60 }
61 
62 template <typename T>
63 static void createPrivateVarSyms(Fortran::lower::AbstractConverter &converter,
64                                  const T *clause) {
65   const Fortran::parser::OmpObjectList &ompObjectList = clause->v;
66   for (const Fortran::parser::OmpObject &ompObject : ompObjectList.v) {
67     Fortran::semantics::Symbol *sym = getOmpObjectSymbol(ompObject);
68     // Privatization for symbols which are pre-determined (like loop index
69     // variables) happen separately, for everything else privatize here.
70     if (sym->test(Fortran::semantics::Symbol::Flag::OmpPreDetermined))
71       continue;
72     bool success = converter.createHostAssociateVarClone(*sym);
73     (void)success;
74     assert(success && "Privatization failed due to existing binding");
75     if constexpr (std::is_same_v<T, Fortran::parser::OmpClause::Firstprivate>) {
76       converter.copyHostAssociateVar(*sym);
77     }
78   }
79 }
80 
81 static void privatizeVars(Fortran::lower::AbstractConverter &converter,
82                           const Fortran::parser::OmpClauseList &opClauseList) {
83   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
84   auto insPt = firOpBuilder.saveInsertionPoint();
85   firOpBuilder.setInsertionPointToStart(firOpBuilder.getAllocaBlock());
86   bool hasFirstPrivateOp = false;
87   for (const Fortran::parser::OmpClause &clause : opClauseList.v) {
88     if (const auto &privateClause =
89             std::get_if<Fortran::parser::OmpClause::Private>(&clause.u)) {
90       createPrivateVarSyms(converter, privateClause);
91     } else if (const auto &firstPrivateClause =
92                    std::get_if<Fortran::parser::OmpClause::Firstprivate>(
93                        &clause.u)) {
94       createPrivateVarSyms(converter, firstPrivateClause);
95       hasFirstPrivateOp = true;
96     }
97   }
98   if (hasFirstPrivateOp)
99     firOpBuilder.create<mlir::omp::BarrierOp>(converter.getCurrentLocation());
100   firOpBuilder.restoreInsertionPoint(insPt);
101 }
102 
103 /// The COMMON block is a global structure. \p commonValue is the base address
104 /// of the the COMMON block. As the offset from the symbol \p sym, generate the
105 /// COMMON block member value (commonValue + offset) for the symbol.
106 /// FIXME: Share the code with `instantiateCommon` in ConvertVariable.cpp.
107 static mlir::Value
108 genCommonBlockMember(Fortran::lower::AbstractConverter &converter,
109                      const Fortran::semantics::Symbol &sym,
110                      mlir::Value commonValue) {
111   auto &firOpBuilder = converter.getFirOpBuilder();
112   mlir::Location currentLocation = converter.getCurrentLocation();
113   mlir::IntegerType i8Ty = firOpBuilder.getIntegerType(8);
114   mlir::Type i8Ptr = firOpBuilder.getRefType(i8Ty);
115   mlir::Type seqTy = firOpBuilder.getRefType(firOpBuilder.getVarLenSeqTy(i8Ty));
116   mlir::Value base =
117       firOpBuilder.createConvert(currentLocation, seqTy, commonValue);
118   std::size_t byteOffset = sym.GetUltimate().offset();
119   mlir::Value offs = firOpBuilder.createIntegerConstant(
120       currentLocation, firOpBuilder.getIndexType(), byteOffset);
121   mlir::Value varAddr = firOpBuilder.create<fir::CoordinateOp>(
122       currentLocation, i8Ptr, base, mlir::ValueRange{offs});
123   mlir::Type symType = converter.genType(sym);
124   return firOpBuilder.createConvert(currentLocation,
125                                     firOpBuilder.getRefType(symType), varAddr);
126 }
127 
128 // Get the extended value for \p val by extracting additional variable
129 // information from \p base.
130 static fir::ExtendedValue getExtendedValue(fir::ExtendedValue base,
131                                            mlir::Value val) {
132   return base.match(
133       [&](const fir::MutableBoxValue &box) -> fir::ExtendedValue {
134         return fir::MutableBoxValue(val, box.nonDeferredLenParams(), {});
135       },
136       [&](const auto &) -> fir::ExtendedValue {
137         return fir::substBase(base, val);
138       });
139 }
140 
141 static void threadPrivatizeVars(Fortran::lower::AbstractConverter &converter,
142                                 Fortran::lower::pft::Evaluation &eval) {
143   auto &firOpBuilder = converter.getFirOpBuilder();
144   mlir::Location currentLocation = converter.getCurrentLocation();
145   auto insPt = firOpBuilder.saveInsertionPoint();
146   firOpBuilder.setInsertionPointToStart(firOpBuilder.getAllocaBlock());
147 
148   // Get the original ThreadprivateOp corresponding to the symbol and use the
149   // symbol value from that opeartion to create one ThreadprivateOp copy
150   // operation inside the parallel region.
151   auto genThreadprivateOp = [&](Fortran::lower::SymbolRef sym) -> mlir::Value {
152     mlir::Value symOriThreadprivateValue = converter.getSymbolAddress(sym);
153     mlir::Operation *op = symOriThreadprivateValue.getDefiningOp();
154     assert(mlir::isa<mlir::omp::ThreadprivateOp>(op) &&
155            "The threadprivate operation not created");
156     mlir::Value symValue =
157         mlir::dyn_cast<mlir::omp::ThreadprivateOp>(op).sym_addr();
158     return firOpBuilder.create<mlir::omp::ThreadprivateOp>(
159         currentLocation, symValue.getType(), symValue);
160   };
161 
162   llvm::SetVector<const Fortran::semantics::Symbol *> threadprivateSyms;
163   converter.collectSymbolSet(eval, threadprivateSyms,
164                              Fortran::semantics::Symbol::Flag::OmpThreadprivate,
165                              /*isUltimateSymbol=*/false);
166   std::set<Fortran::semantics::SourceName> threadprivateSymNames;
167 
168   // For a COMMON block, the ThreadprivateOp is generated for itself instead of
169   // its members, so only bind the value of the new copied ThreadprivateOp
170   // inside the parallel region to the common block symbol only once for
171   // multiple members in one COMMON block.
172   llvm::SetVector<const Fortran::semantics::Symbol *> commonSyms;
173   for (std::size_t i = 0; i < threadprivateSyms.size(); i++) {
174     auto sym = threadprivateSyms[i];
175     mlir::Value symThreadprivateValue;
176     // The variable may be used more than once, and each reference has one
177     // symbol with the same name. Only do once for references of one variable.
178     if (threadprivateSymNames.find(sym->name()) != threadprivateSymNames.end())
179       continue;
180     threadprivateSymNames.insert(sym->name());
181     if (const Fortran::semantics::Symbol *common =
182             Fortran::semantics::FindCommonBlockContaining(sym->GetUltimate())) {
183       mlir::Value commonThreadprivateValue;
184       if (commonSyms.contains(common)) {
185         commonThreadprivateValue = converter.getSymbolAddress(*common);
186       } else {
187         commonThreadprivateValue = genThreadprivateOp(*common);
188         converter.bindSymbol(*common, commonThreadprivateValue);
189         commonSyms.insert(common);
190       }
191       symThreadprivateValue =
192           genCommonBlockMember(converter, *sym, commonThreadprivateValue);
193     } else {
194       symThreadprivateValue = genThreadprivateOp(*sym);
195     }
196 
197     fir::ExtendedValue sexv = converter.getSymbolExtendedValue(*sym);
198     fir::ExtendedValue symThreadprivateExv =
199         getExtendedValue(sexv, symThreadprivateValue);
200     converter.bindSymbol(*sym, symThreadprivateExv);
201   }
202 
203   firOpBuilder.restoreInsertionPoint(insPt);
204 }
205 
206 static void
207 genCopyinClause(Fortran::lower::AbstractConverter &converter,
208                 const Fortran::parser::OmpClauseList &opClauseList) {
209   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
210   mlir::OpBuilder::InsertPoint insPt = firOpBuilder.saveInsertionPoint();
211   firOpBuilder.setInsertionPointToStart(firOpBuilder.getAllocaBlock());
212   bool hasCopyin = false;
213   for (const Fortran::parser::OmpClause &clause : opClauseList.v) {
214     if (const auto &copyinClause =
215             std::get_if<Fortran::parser::OmpClause::Copyin>(&clause.u)) {
216       hasCopyin = true;
217       const Fortran::parser::OmpObjectList &ompObjectList = copyinClause->v;
218       for (const Fortran::parser::OmpObject &ompObject : ompObjectList.v) {
219         Fortran::semantics::Symbol *sym = getOmpObjectSymbol(ompObject);
220         if (sym->has<Fortran::semantics::CommonBlockDetails>())
221           TODO(converter.getCurrentLocation(), "common block in Copyin clause");
222         if (Fortran::semantics::IsAllocatableOrPointer(sym->GetUltimate()))
223           TODO(converter.getCurrentLocation(),
224                "pointer or allocatable variables in Copyin clause");
225         assert(sym->has<Fortran::semantics::HostAssocDetails>() &&
226                "No host-association found");
227         converter.copyHostAssociateVar(*sym);
228       }
229     }
230   }
231   if (hasCopyin)
232     firOpBuilder.create<mlir::omp::BarrierOp>(converter.getCurrentLocation());
233   firOpBuilder.restoreInsertionPoint(insPt);
234 }
235 
236 static void genObjectList(const Fortran::parser::OmpObjectList &objectList,
237                           Fortran::lower::AbstractConverter &converter,
238                           llvm::SmallVectorImpl<Value> &operands) {
239   auto addOperands = [&](Fortran::lower::SymbolRef sym) {
240     const mlir::Value variable = converter.getSymbolAddress(sym);
241     if (variable) {
242       operands.push_back(variable);
243     } else {
244       if (const auto *details =
245               sym->detailsIf<Fortran::semantics::HostAssocDetails>()) {
246         operands.push_back(converter.getSymbolAddress(details->symbol()));
247         converter.copySymbolBinding(details->symbol(), sym);
248       }
249     }
250   };
251   for (const Fortran::parser::OmpObject &ompObject : objectList.v) {
252     Fortran::semantics::Symbol *sym = getOmpObjectSymbol(ompObject);
253     addOperands(*sym);
254   }
255 }
256 
257 static mlir::Type getLoopVarType(Fortran::lower::AbstractConverter &converter,
258                                  std::size_t loopVarTypeSize) {
259   // OpenMP runtime requires 32-bit or 64-bit loop variables.
260   loopVarTypeSize = loopVarTypeSize * 8;
261   if (loopVarTypeSize < 32) {
262     loopVarTypeSize = 32;
263   } else if (loopVarTypeSize > 64) {
264     loopVarTypeSize = 64;
265     mlir::emitWarning(converter.getCurrentLocation(),
266                       "OpenMP loop iteration variable cannot have more than 64 "
267                       "bits size and will be narrowed into 64 bits.");
268   }
269   assert((loopVarTypeSize == 32 || loopVarTypeSize == 64) &&
270          "OpenMP loop iteration variable size must be transformed into 32-bit "
271          "or 64-bit");
272   return converter.getFirOpBuilder().getIntegerType(loopVarTypeSize);
273 }
274 
275 /// Create empty blocks for the current region.
276 /// These blocks replace blocks parented to an enclosing region.
277 void createEmptyRegionBlocks(
278     fir::FirOpBuilder &firOpBuilder,
279     std::list<Fortran::lower::pft::Evaluation> &evaluationList) {
280   auto *region = &firOpBuilder.getRegion();
281   for (auto &eval : evaluationList) {
282     if (eval.block) {
283       if (eval.block->empty()) {
284         eval.block->erase();
285         eval.block = firOpBuilder.createBlock(region);
286       } else {
287         [[maybe_unused]] auto &terminatorOp = eval.block->back();
288         assert((mlir::isa<mlir::omp::TerminatorOp>(terminatorOp) ||
289                 mlir::isa<mlir::omp::YieldOp>(terminatorOp)) &&
290                "expected terminator op");
291       }
292     }
293     if (!eval.isDirective() && eval.hasNestedEvaluations())
294       createEmptyRegionBlocks(firOpBuilder, eval.getNestedEvaluations());
295   }
296 }
297 
298 /// Create the body (block) for an OpenMP Operation.
299 ///
300 /// \param [in]    op - the operation the body belongs to.
301 /// \param [inout] converter - converter to use for the clauses.
302 /// \param [in]    loc - location in source code.
303 /// \param [in]    eval - current PFT node/evaluation.
304 /// \oaran [in]    clauses - list of clauses to process.
305 /// \param [in]    args - block arguments (induction variable[s]) for the
306 ////                      region.
307 /// \param [in]    outerCombined - is this an outer operation - prevents
308 ///                                privatization.
309 template <typename Op>
310 static void
311 createBodyOfOp(Op &op, Fortran::lower::AbstractConverter &converter,
312                mlir::Location &loc, Fortran::lower::pft::Evaluation &eval,
313                const Fortran::parser::OmpClauseList *clauses = nullptr,
314                const SmallVector<const Fortran::semantics::Symbol *> &args = {},
315                bool outerCombined = false) {
316   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
317   // If an argument for the region is provided then create the block with that
318   // argument. Also update the symbol's address with the mlir argument value.
319   // e.g. For loops the argument is the induction variable. And all further
320   // uses of the induction variable should use this mlir value.
321   mlir::Operation *storeOp = nullptr;
322   if (args.size()) {
323     std::size_t loopVarTypeSize = 0;
324     for (const Fortran::semantics::Symbol *arg : args)
325       loopVarTypeSize = std::max(loopVarTypeSize, arg->GetUltimate().size());
326     mlir::Type loopVarType = getLoopVarType(converter, loopVarTypeSize);
327     SmallVector<Type> tiv;
328     SmallVector<Location> locs;
329     for (int i = 0; i < (int)args.size(); i++) {
330       tiv.push_back(loopVarType);
331       locs.push_back(loc);
332     }
333     firOpBuilder.createBlock(&op.getRegion(), {}, tiv, locs);
334     int argIndex = 0;
335     // The argument is not currently in memory, so make a temporary for the
336     // argument, and store it there, then bind that location to the argument.
337     for (const Fortran::semantics::Symbol *arg : args) {
338       mlir::Value val =
339           fir::getBase(op.getRegion().front().getArgument(argIndex));
340       mlir::Value temp = firOpBuilder.createTemporary(
341           loc, loopVarType,
342           llvm::ArrayRef<mlir::NamedAttribute>{
343               Fortran::lower::getAdaptToByRefAttr(firOpBuilder)});
344       storeOp = firOpBuilder.create<fir::StoreOp>(loc, val, temp);
345       converter.bindSymbol(*arg, temp);
346       argIndex++;
347     }
348   } else {
349     firOpBuilder.createBlock(&op.getRegion());
350   }
351   // Set the insert for the terminator operation to go at the end of the
352   // block - this is either empty or the block with the stores above,
353   // the end of the block works for both.
354   mlir::Block &block = op.getRegion().back();
355   firOpBuilder.setInsertionPointToEnd(&block);
356 
357   // If it is an unstructured region and is not the outer region of a combined
358   // construct, create empty blocks for all evaluations.
359   if (eval.lowerAsUnstructured() && !outerCombined)
360     createEmptyRegionBlocks(firOpBuilder, eval.getNestedEvaluations());
361 
362   // Insert the terminator.
363   if constexpr (std::is_same_v<Op, omp::WsLoopOp> ||
364                 std::is_same_v<Op, omp::SimdLoopOp>) {
365     mlir::ValueRange results;
366     firOpBuilder.create<mlir::omp::YieldOp>(loc, results);
367   } else {
368     firOpBuilder.create<mlir::omp::TerminatorOp>(loc);
369   }
370 
371   // Reset the insert point to before the terminator.
372   if (storeOp)
373     firOpBuilder.setInsertionPointAfter(storeOp);
374   else
375     firOpBuilder.setInsertionPointToStart(&block);
376 
377   // Handle privatization. Do not privatize if this is the outer operation.
378   if (clauses && !outerCombined)
379     privatizeVars(converter, *clauses);
380 
381   if (std::is_same_v<Op, omp::ParallelOp>) {
382     threadPrivatizeVars(converter, eval);
383     if (clauses)
384       genCopyinClause(converter, *clauses);
385   }
386 }
387 
388 static void genOMP(Fortran::lower::AbstractConverter &converter,
389                    Fortran::lower::pft::Evaluation &eval,
390                    const Fortran::parser::OpenMPSimpleStandaloneConstruct
391                        &simpleStandaloneConstruct) {
392   const auto &directive =
393       std::get<Fortran::parser::OmpSimpleStandaloneDirective>(
394           simpleStandaloneConstruct.t);
395   switch (directive.v) {
396   default:
397     break;
398   case llvm::omp::Directive::OMPD_barrier:
399     converter.getFirOpBuilder().create<mlir::omp::BarrierOp>(
400         converter.getCurrentLocation());
401     break;
402   case llvm::omp::Directive::OMPD_taskwait:
403     converter.getFirOpBuilder().create<mlir::omp::TaskwaitOp>(
404         converter.getCurrentLocation());
405     break;
406   case llvm::omp::Directive::OMPD_taskyield:
407     converter.getFirOpBuilder().create<mlir::omp::TaskyieldOp>(
408         converter.getCurrentLocation());
409     break;
410   case llvm::omp::Directive::OMPD_target_enter_data:
411     TODO(converter.getCurrentLocation(), "OMPD_target_enter_data");
412   case llvm::omp::Directive::OMPD_target_exit_data:
413     TODO(converter.getCurrentLocation(), "OMPD_target_exit_data");
414   case llvm::omp::Directive::OMPD_target_update:
415     TODO(converter.getCurrentLocation(), "OMPD_target_update");
416   case llvm::omp::Directive::OMPD_ordered:
417     TODO(converter.getCurrentLocation(), "OMPD_ordered");
418   }
419 }
420 
421 static void
422 genAllocateClause(Fortran::lower::AbstractConverter &converter,
423                   const Fortran::parser::OmpAllocateClause &ompAllocateClause,
424                   SmallVector<Value> &allocatorOperands,
425                   SmallVector<Value> &allocateOperands) {
426   auto &firOpBuilder = converter.getFirOpBuilder();
427   auto currentLocation = converter.getCurrentLocation();
428   Fortran::lower::StatementContext stmtCtx;
429 
430   mlir::Value allocatorOperand;
431   const Fortran::parser::OmpObjectList &ompObjectList =
432       std::get<Fortran::parser::OmpObjectList>(ompAllocateClause.t);
433   const auto &allocatorValue =
434       std::get<std::optional<Fortran::parser::OmpAllocateClause::Allocator>>(
435           ompAllocateClause.t);
436   // Check if allocate clause has allocator specified. If so, add it
437   // to list of allocators, otherwise, add default allocator to
438   // list of allocators.
439   if (allocatorValue) {
440     allocatorOperand = fir::getBase(converter.genExprValue(
441         *Fortran::semantics::GetExpr(allocatorValue->v), stmtCtx));
442     allocatorOperands.insert(allocatorOperands.end(), ompObjectList.v.size(),
443                              allocatorOperand);
444   } else {
445     allocatorOperand = firOpBuilder.createIntegerConstant(
446         currentLocation, firOpBuilder.getI32Type(), 1);
447     allocatorOperands.insert(allocatorOperands.end(), ompObjectList.v.size(),
448                              allocatorOperand);
449   }
450   genObjectList(ompObjectList, converter, allocateOperands);
451 }
452 
453 static void
454 genOMP(Fortran::lower::AbstractConverter &converter,
455        Fortran::lower::pft::Evaluation &eval,
456        const Fortran::parser::OpenMPStandaloneConstruct &standaloneConstruct) {
457   std::visit(
458       Fortran::common::visitors{
459           [&](const Fortran::parser::OpenMPSimpleStandaloneConstruct
460                   &simpleStandaloneConstruct) {
461             genOMP(converter, eval, simpleStandaloneConstruct);
462           },
463           [&](const Fortran::parser::OpenMPFlushConstruct &flushConstruct) {
464             SmallVector<Value, 4> operandRange;
465             if (const auto &ompObjectList =
466                     std::get<std::optional<Fortran::parser::OmpObjectList>>(
467                         flushConstruct.t))
468               genObjectList(*ompObjectList, converter, operandRange);
469             const auto &memOrderClause = std::get<std::optional<
470                 std::list<Fortran::parser::OmpMemoryOrderClause>>>(
471                 flushConstruct.t);
472             if (memOrderClause.has_value() && memOrderClause->size() > 0)
473               TODO(converter.getCurrentLocation(),
474                    "Handle OmpMemoryOrderClause");
475             converter.getFirOpBuilder().create<mlir::omp::FlushOp>(
476                 converter.getCurrentLocation(), operandRange);
477           },
478           [&](const Fortran::parser::OpenMPCancelConstruct &cancelConstruct) {
479             TODO(converter.getCurrentLocation(), "OpenMPCancelConstruct");
480           },
481           [&](const Fortran::parser::OpenMPCancellationPointConstruct
482                   &cancellationPointConstruct) {
483             TODO(converter.getCurrentLocation(), "OpenMPCancelConstruct");
484           },
485       },
486       standaloneConstruct.u);
487 }
488 
489 static omp::ClauseProcBindKindAttr genProcBindKindAttr(
490     fir::FirOpBuilder &firOpBuilder,
491     const Fortran::parser::OmpClause::ProcBind *procBindClause) {
492   omp::ClauseProcBindKind pbKind;
493   switch (procBindClause->v.v) {
494   case Fortran::parser::OmpProcBindClause::Type::Master:
495     pbKind = omp::ClauseProcBindKind::Master;
496     break;
497   case Fortran::parser::OmpProcBindClause::Type::Close:
498     pbKind = omp::ClauseProcBindKind::Close;
499     break;
500   case Fortran::parser::OmpProcBindClause::Type::Spread:
501     pbKind = omp::ClauseProcBindKind::Spread;
502     break;
503   case Fortran::parser::OmpProcBindClause::Type::Primary:
504     pbKind = omp::ClauseProcBindKind::Primary;
505     break;
506   }
507   return omp::ClauseProcBindKindAttr::get(firOpBuilder.getContext(), pbKind);
508 }
509 
510 /* When parallel is used in a combined construct, then use this function to
511  * create the parallel operation. It handles the parallel specific clauses
512  * and leaves the rest for handling at the inner operations.
513  * TODO: Refactor clause handling
514  */
515 template <typename Directive>
516 static void
517 createCombinedParallelOp(Fortran::lower::AbstractConverter &converter,
518                          Fortran::lower::pft::Evaluation &eval,
519                          const Directive &directive) {
520   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
521   mlir::Location currentLocation = converter.getCurrentLocation();
522   Fortran::lower::StatementContext stmtCtx;
523   llvm::ArrayRef<mlir::Type> argTy;
524   mlir::Value ifClauseOperand, numThreadsClauseOperand;
525   SmallVector<Value> allocatorOperands, allocateOperands;
526   mlir::omp::ClauseProcBindKindAttr procBindKindAttr;
527   const auto &opClauseList =
528       std::get<Fortran::parser::OmpClauseList>(directive.t);
529   // TODO: Handle the following clauses
530   // 1. default
531   // Note: rest of the clauses are handled when the inner operation is created
532   for (const Fortran::parser::OmpClause &clause : opClauseList.v) {
533     if (const auto &ifClause =
534             std::get_if<Fortran::parser::OmpClause::If>(&clause.u)) {
535       auto &expr = std::get<Fortran::parser::ScalarLogicalExpr>(ifClause->v.t);
536       mlir::Value ifVal = fir::getBase(
537           converter.genExprValue(*Fortran::semantics::GetExpr(expr), stmtCtx));
538       ifClauseOperand = firOpBuilder.createConvert(
539           currentLocation, firOpBuilder.getI1Type(), ifVal);
540     } else if (const auto &numThreadsClause =
541                    std::get_if<Fortran::parser::OmpClause::NumThreads>(
542                        &clause.u)) {
543       numThreadsClauseOperand = fir::getBase(converter.genExprValue(
544           *Fortran::semantics::GetExpr(numThreadsClause->v), stmtCtx));
545     } else if (const auto &procBindClause =
546                    std::get_if<Fortran::parser::OmpClause::ProcBind>(
547                        &clause.u)) {
548       procBindKindAttr = genProcBindKindAttr(firOpBuilder, procBindClause);
549     }
550   }
551   // Create and insert the operation.
552   auto parallelOp = firOpBuilder.create<mlir::omp::ParallelOp>(
553       currentLocation, argTy, ifClauseOperand, numThreadsClauseOperand,
554       allocateOperands, allocatorOperands, /*reduction_vars=*/ValueRange(),
555       /*reductions=*/nullptr, procBindKindAttr);
556 
557   createBodyOfOp<omp::ParallelOp>(parallelOp, converter, currentLocation, eval,
558                                   &opClauseList, /*iv=*/{},
559                                   /*isCombined=*/true);
560 }
561 
562 static void
563 genOMP(Fortran::lower::AbstractConverter &converter,
564        Fortran::lower::pft::Evaluation &eval,
565        const Fortran::parser::OpenMPBlockConstruct &blockConstruct) {
566   const auto &beginBlockDirective =
567       std::get<Fortran::parser::OmpBeginBlockDirective>(blockConstruct.t);
568   const auto &blockDirective =
569       std::get<Fortran::parser::OmpBlockDirective>(beginBlockDirective.t);
570   const auto &endBlockDirective =
571       std::get<Fortran::parser::OmpEndBlockDirective>(blockConstruct.t);
572   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
573   mlir::Location currentLocation = converter.getCurrentLocation();
574 
575   Fortran::lower::StatementContext stmtCtx;
576   llvm::ArrayRef<mlir::Type> argTy;
577   mlir::Value ifClauseOperand, numThreadsClauseOperand, finalClauseOperand,
578       priorityClauseOperand;
579   mlir::omp::ClauseProcBindKindAttr procBindKindAttr;
580   SmallVector<Value> allocateOperands, allocatorOperands;
581   mlir::UnitAttr nowaitAttr, untiedAttr, mergeableAttr;
582 
583   const auto &opClauseList =
584       std::get<Fortran::parser::OmpClauseList>(beginBlockDirective.t);
585   for (const auto &clause : opClauseList.v) {
586     if (const auto &ifClause =
587             std::get_if<Fortran::parser::OmpClause::If>(&clause.u)) {
588       auto &expr = std::get<Fortran::parser::ScalarLogicalExpr>(ifClause->v.t);
589       mlir::Value ifVal = fir::getBase(
590           converter.genExprValue(*Fortran::semantics::GetExpr(expr), stmtCtx));
591       ifClauseOperand = firOpBuilder.createConvert(
592           currentLocation, firOpBuilder.getI1Type(), ifVal);
593     } else if (const auto &numThreadsClause =
594                    std::get_if<Fortran::parser::OmpClause::NumThreads>(
595                        &clause.u)) {
596       // OMPIRBuilder expects `NUM_THREAD` clause as a `Value`.
597       numThreadsClauseOperand = fir::getBase(converter.genExprValue(
598           *Fortran::semantics::GetExpr(numThreadsClause->v), stmtCtx));
599     } else if (const auto &procBindClause =
600                    std::get_if<Fortran::parser::OmpClause::ProcBind>(
601                        &clause.u)) {
602       procBindKindAttr = genProcBindKindAttr(firOpBuilder, procBindClause);
603     } else if (const auto &allocateClause =
604                    std::get_if<Fortran::parser::OmpClause::Allocate>(
605                        &clause.u)) {
606       genAllocateClause(converter, allocateClause->v, allocatorOperands,
607                         allocateOperands);
608     } else if (std::get_if<Fortran::parser::OmpClause::Private>(&clause.u) ||
609                std::get_if<Fortran::parser::OmpClause::Firstprivate>(
610                    &clause.u) ||
611                std::get_if<Fortran::parser::OmpClause::Copyin>(&clause.u)) {
612       // Privatisation and copyin clauses are handled elsewhere.
613       continue;
614     } else if (std::get_if<Fortran::parser::OmpClause::Threads>(&clause.u)) {
615       // Nothing needs to be done for threads clause.
616       continue;
617     } else if (const auto &finalClause =
618                    std::get_if<Fortran::parser::OmpClause::Final>(&clause.u)) {
619       mlir::Value finalVal = fir::getBase(converter.genExprValue(
620           *Fortran::semantics::GetExpr(finalClause->v), stmtCtx));
621       finalClauseOperand = firOpBuilder.createConvert(
622           currentLocation, firOpBuilder.getI1Type(), finalVal);
623     } else if (std::get_if<Fortran::parser::OmpClause::Untied>(&clause.u)) {
624       untiedAttr = firOpBuilder.getUnitAttr();
625     } else if (std::get_if<Fortran::parser::OmpClause::Mergeable>(&clause.u)) {
626       mergeableAttr = firOpBuilder.getUnitAttr();
627     } else if (const auto &priorityClause =
628                    std::get_if<Fortran::parser::OmpClause::Priority>(
629                        &clause.u)) {
630       priorityClauseOperand = fir::getBase(converter.genExprValue(
631           *Fortran::semantics::GetExpr(priorityClause->v), stmtCtx));
632     } else {
633       TODO(currentLocation, "OpenMP Block construct clauses");
634     }
635   }
636 
637   for (const auto &clause :
638        std::get<Fortran::parser::OmpClauseList>(endBlockDirective.t).v) {
639     if (std::get_if<Fortran::parser::OmpClause::Nowait>(&clause.u))
640       nowaitAttr = firOpBuilder.getUnitAttr();
641   }
642 
643   if (blockDirective.v == llvm::omp::OMPD_parallel) {
644     // Create and insert the operation.
645     auto parallelOp = firOpBuilder.create<mlir::omp::ParallelOp>(
646         currentLocation, argTy, ifClauseOperand, numThreadsClauseOperand,
647         allocateOperands, allocatorOperands, /*reduction_vars=*/ValueRange(),
648         /*reductions=*/nullptr, procBindKindAttr);
649     createBodyOfOp<omp::ParallelOp>(parallelOp, converter, currentLocation,
650                                     eval, &opClauseList);
651   } else if (blockDirective.v == llvm::omp::OMPD_master) {
652     auto masterOp =
653         firOpBuilder.create<mlir::omp::MasterOp>(currentLocation, argTy);
654     createBodyOfOp<omp::MasterOp>(masterOp, converter, currentLocation, eval);
655   } else if (blockDirective.v == llvm::omp::OMPD_single) {
656     auto singleOp = firOpBuilder.create<mlir::omp::SingleOp>(
657         currentLocation, allocateOperands, allocatorOperands, nowaitAttr);
658     createBodyOfOp<omp::SingleOp>(singleOp, converter, currentLocation, eval);
659   } else if (blockDirective.v == llvm::omp::OMPD_ordered) {
660     auto orderedOp = firOpBuilder.create<mlir::omp::OrderedRegionOp>(
661         currentLocation, /*simd=*/nullptr);
662     createBodyOfOp<omp::OrderedRegionOp>(orderedOp, converter, currentLocation,
663                                          eval);
664   } else if (blockDirective.v == llvm::omp::OMPD_task) {
665     auto taskOp = firOpBuilder.create<mlir::omp::TaskOp>(
666         currentLocation, ifClauseOperand, finalClauseOperand, untiedAttr,
667         mergeableAttr, /*in_reduction_vars=*/ValueRange(),
668         /*in_reductions=*/nullptr, priorityClauseOperand, allocateOperands,
669         allocatorOperands);
670     createBodyOfOp(taskOp, converter, currentLocation, eval, &opClauseList);
671   } else {
672     TODO(converter.getCurrentLocation(), "Unhandled block directive");
673   }
674 }
675 
676 static mlir::omp::ScheduleModifier
677 translateModifier(const Fortran::parser::OmpScheduleModifierType &m) {
678   switch (m.v) {
679   case Fortran::parser::OmpScheduleModifierType::ModType::Monotonic:
680     return mlir::omp::ScheduleModifier::monotonic;
681   case Fortran::parser::OmpScheduleModifierType::ModType::Nonmonotonic:
682     return mlir::omp::ScheduleModifier::nonmonotonic;
683   case Fortran::parser::OmpScheduleModifierType::ModType::Simd:
684     return mlir::omp::ScheduleModifier::simd;
685   }
686   return mlir::omp::ScheduleModifier::none;
687 }
688 
689 static mlir::omp::ScheduleModifier
690 getScheduleModifier(const Fortran::parser::OmpScheduleClause &x) {
691   const auto &modifier =
692       std::get<std::optional<Fortran::parser::OmpScheduleModifier>>(x.t);
693   // The input may have the modifier any order, so we look for one that isn't
694   // SIMD. If modifier is not set at all, fall down to the bottom and return
695   // "none".
696   if (modifier) {
697     const auto &modType1 =
698         std::get<Fortran::parser::OmpScheduleModifier::Modifier1>(modifier->t);
699     if (modType1.v.v ==
700         Fortran::parser::OmpScheduleModifierType::ModType::Simd) {
701       const auto &modType2 = std::get<
702           std::optional<Fortran::parser::OmpScheduleModifier::Modifier2>>(
703           modifier->t);
704       if (modType2 &&
705           modType2->v.v !=
706               Fortran::parser::OmpScheduleModifierType::ModType::Simd)
707         return translateModifier(modType2->v);
708 
709       return mlir::omp::ScheduleModifier::none;
710     }
711 
712     return translateModifier(modType1.v);
713   }
714   return mlir::omp::ScheduleModifier::none;
715 }
716 
717 static mlir::omp::ScheduleModifier
718 getSIMDModifier(const Fortran::parser::OmpScheduleClause &x) {
719   const auto &modifier =
720       std::get<std::optional<Fortran::parser::OmpScheduleModifier>>(x.t);
721   // Either of the two possible modifiers in the input can be the SIMD modifier,
722   // so look in either one, and return simd if we find one. Not found = return
723   // "none".
724   if (modifier) {
725     const auto &modType1 =
726         std::get<Fortran::parser::OmpScheduleModifier::Modifier1>(modifier->t);
727     if (modType1.v.v == Fortran::parser::OmpScheduleModifierType::ModType::Simd)
728       return mlir::omp::ScheduleModifier::simd;
729 
730     const auto &modType2 = std::get<
731         std::optional<Fortran::parser::OmpScheduleModifier::Modifier2>>(
732         modifier->t);
733     if (modType2 && modType2->v.v ==
734                         Fortran::parser::OmpScheduleModifierType::ModType::Simd)
735       return mlir::omp::ScheduleModifier::simd;
736   }
737   return mlir::omp::ScheduleModifier::none;
738 }
739 
740 static void genOMP(Fortran::lower::AbstractConverter &converter,
741                    Fortran::lower::pft::Evaluation &eval,
742                    const Fortran::parser::OpenMPLoopConstruct &loopConstruct) {
743 
744   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
745   mlir::Location currentLocation = converter.getCurrentLocation();
746   llvm::SmallVector<mlir::Value> lowerBound, upperBound, step, linearVars,
747       linearStepVars, reductionVars;
748   mlir::Value scheduleChunkClauseOperand;
749   mlir::Attribute scheduleClauseOperand, collapseClauseOperand,
750       noWaitClauseOperand, orderedClauseOperand, orderClauseOperand;
751   const auto &loopOpClauseList = std::get<Fortran::parser::OmpClauseList>(
752       std::get<Fortran::parser::OmpBeginLoopDirective>(loopConstruct.t).t);
753 
754   const auto ompDirective =
755       std::get<Fortran::parser::OmpLoopDirective>(
756           std::get<Fortran::parser::OmpBeginLoopDirective>(loopConstruct.t).t)
757           .v;
758   if (llvm::omp::OMPD_parallel_do == ompDirective) {
759     createCombinedParallelOp<Fortran::parser::OmpBeginLoopDirective>(
760         converter, eval,
761         std::get<Fortran::parser::OmpBeginLoopDirective>(loopConstruct.t));
762   } else if (llvm::omp::OMPD_do != ompDirective &&
763              llvm::omp::OMPD_simd != ompDirective) {
764     TODO(converter.getCurrentLocation(), "Construct enclosing do loop");
765   }
766 
767   // Collect the loops to collapse.
768   auto *doConstructEval = &eval.getFirstNestedEvaluation();
769 
770   std::int64_t collapseValue =
771       Fortran::lower::getCollapseValue(loopOpClauseList);
772   std::size_t loopVarTypeSize = 0;
773   SmallVector<const Fortran::semantics::Symbol *> iv;
774   do {
775     auto *doLoop = &doConstructEval->getFirstNestedEvaluation();
776     auto *doStmt = doLoop->getIf<Fortran::parser::NonLabelDoStmt>();
777     assert(doStmt && "Expected do loop to be in the nested evaluation");
778     const auto &loopControl =
779         std::get<std::optional<Fortran::parser::LoopControl>>(doStmt->t);
780     const Fortran::parser::LoopControl::Bounds *bounds =
781         std::get_if<Fortran::parser::LoopControl::Bounds>(&loopControl->u);
782     assert(bounds && "Expected bounds for worksharing do loop");
783     Fortran::lower::StatementContext stmtCtx;
784     lowerBound.push_back(fir::getBase(converter.genExprValue(
785         *Fortran::semantics::GetExpr(bounds->lower), stmtCtx)));
786     upperBound.push_back(fir::getBase(converter.genExprValue(
787         *Fortran::semantics::GetExpr(bounds->upper), stmtCtx)));
788     if (bounds->step) {
789       step.push_back(fir::getBase(converter.genExprValue(
790           *Fortran::semantics::GetExpr(bounds->step), stmtCtx)));
791     } else { // If `step` is not present, assume it as `1`.
792       step.push_back(firOpBuilder.createIntegerConstant(
793           currentLocation, firOpBuilder.getIntegerType(32), 1));
794     }
795     iv.push_back(bounds->name.thing.symbol);
796     loopVarTypeSize = std::max(loopVarTypeSize,
797                                bounds->name.thing.symbol->GetUltimate().size());
798 
799     collapseValue--;
800     doConstructEval =
801         &*std::next(doConstructEval->getNestedEvaluations().begin());
802   } while (collapseValue > 0);
803 
804   for (const auto &clause : loopOpClauseList.v) {
805     if (const auto &scheduleClause =
806             std::get_if<Fortran::parser::OmpClause::Schedule>(&clause.u)) {
807       if (const auto &chunkExpr =
808               std::get<std::optional<Fortran::parser::ScalarIntExpr>>(
809                   scheduleClause->v.t)) {
810         if (const auto *expr = Fortran::semantics::GetExpr(*chunkExpr)) {
811           Fortran::lower::StatementContext stmtCtx;
812           scheduleChunkClauseOperand =
813               fir::getBase(converter.genExprValue(*expr, stmtCtx));
814         }
815       }
816     }
817   }
818 
819   // The types of lower bound, upper bound, and step are converted into the
820   // type of the loop variable if necessary.
821   mlir::Type loopVarType = getLoopVarType(converter, loopVarTypeSize);
822   for (unsigned it = 0; it < (unsigned)lowerBound.size(); it++) {
823     lowerBound[it] = firOpBuilder.createConvert(currentLocation, loopVarType,
824                                                 lowerBound[it]);
825     upperBound[it] = firOpBuilder.createConvert(currentLocation, loopVarType,
826                                                 upperBound[it]);
827     step[it] =
828         firOpBuilder.createConvert(currentLocation, loopVarType, step[it]);
829   }
830 
831   // 2.9.3.1 SIMD construct
832   // TODO: Support all the clauses
833   if (llvm::omp::OMPD_simd == ompDirective) {
834     TypeRange resultType;
835     auto SimdLoopOp = firOpBuilder.create<mlir::omp::SimdLoopOp>(
836         currentLocation, resultType, lowerBound, upperBound, step);
837     createBodyOfOp<omp::SimdLoopOp>(SimdLoopOp, converter, currentLocation,
838                                     eval, &loopOpClauseList, iv);
839     return;
840   }
841 
842   // FIXME: Add support for following clauses:
843   // 1. linear
844   // 2. order
845   auto wsLoopOp = firOpBuilder.create<mlir::omp::WsLoopOp>(
846       currentLocation, lowerBound, upperBound, step, linearVars, linearStepVars,
847       reductionVars, /*reductions=*/nullptr,
848       scheduleClauseOperand.dyn_cast_or_null<omp::ClauseScheduleKindAttr>(),
849       scheduleChunkClauseOperand, /*schedule_modifiers=*/nullptr,
850       /*simd_modifier=*/nullptr,
851       collapseClauseOperand.dyn_cast_or_null<IntegerAttr>(),
852       noWaitClauseOperand.dyn_cast_or_null<UnitAttr>(),
853       orderedClauseOperand.dyn_cast_or_null<IntegerAttr>(),
854       orderClauseOperand.dyn_cast_or_null<omp::ClauseOrderKindAttr>(),
855       /*inclusive=*/firOpBuilder.getUnitAttr());
856 
857   // Handle attribute based clauses.
858   for (const Fortran::parser::OmpClause &clause : loopOpClauseList.v) {
859     if (const auto &orderedClause =
860             std::get_if<Fortran::parser::OmpClause::Ordered>(&clause.u)) {
861       if (orderedClause->v.has_value()) {
862         const auto *expr = Fortran::semantics::GetExpr(orderedClause->v);
863         const std::optional<std::int64_t> orderedClauseValue =
864             Fortran::evaluate::ToInt64(*expr);
865         wsLoopOp.ordered_valAttr(
866             firOpBuilder.getI64IntegerAttr(*orderedClauseValue));
867       } else {
868         wsLoopOp.ordered_valAttr(firOpBuilder.getI64IntegerAttr(0));
869       }
870     } else if (const auto &collapseClause =
871                    std::get_if<Fortran::parser::OmpClause::Collapse>(
872                        &clause.u)) {
873       const auto *expr = Fortran::semantics::GetExpr(collapseClause->v);
874       const std::optional<std::int64_t> collapseValue =
875           Fortran::evaluate::ToInt64(*expr);
876       wsLoopOp.collapse_valAttr(firOpBuilder.getI64IntegerAttr(*collapseValue));
877     } else if (const auto &scheduleClause =
878                    std::get_if<Fortran::parser::OmpClause::Schedule>(
879                        &clause.u)) {
880       mlir::MLIRContext *context = firOpBuilder.getContext();
881       const auto &scheduleType = scheduleClause->v;
882       const auto &scheduleKind =
883           std::get<Fortran::parser::OmpScheduleClause::ScheduleType>(
884               scheduleType.t);
885       switch (scheduleKind) {
886       case Fortran::parser::OmpScheduleClause::ScheduleType::Static:
887         wsLoopOp.schedule_valAttr(omp::ClauseScheduleKindAttr::get(
888             context, omp::ClauseScheduleKind::Static));
889         break;
890       case Fortran::parser::OmpScheduleClause::ScheduleType::Dynamic:
891         wsLoopOp.schedule_valAttr(omp::ClauseScheduleKindAttr::get(
892             context, omp::ClauseScheduleKind::Dynamic));
893         break;
894       case Fortran::parser::OmpScheduleClause::ScheduleType::Guided:
895         wsLoopOp.schedule_valAttr(omp::ClauseScheduleKindAttr::get(
896             context, omp::ClauseScheduleKind::Guided));
897         break;
898       case Fortran::parser::OmpScheduleClause::ScheduleType::Auto:
899         wsLoopOp.schedule_valAttr(omp::ClauseScheduleKindAttr::get(
900             context, omp::ClauseScheduleKind::Auto));
901         break;
902       case Fortran::parser::OmpScheduleClause::ScheduleType::Runtime:
903         wsLoopOp.schedule_valAttr(omp::ClauseScheduleKindAttr::get(
904             context, omp::ClauseScheduleKind::Runtime));
905         break;
906       }
907       mlir::omp::ScheduleModifier scheduleModifier =
908           getScheduleModifier(scheduleClause->v);
909       if (scheduleModifier != mlir::omp::ScheduleModifier::none)
910         wsLoopOp.schedule_modifierAttr(
911             omp::ScheduleModifierAttr::get(context, scheduleModifier));
912       if (getSIMDModifier(scheduleClause->v) !=
913           mlir::omp::ScheduleModifier::none)
914         wsLoopOp.simd_modifierAttr(firOpBuilder.getUnitAttr());
915     }
916   }
917   // In FORTRAN `nowait` clause occur at the end of `omp do` directive.
918   // i.e
919   // !$omp do
920   // <...>
921   // !$omp end do nowait
922   if (const auto &endClauseList =
923           std::get<std::optional<Fortran::parser::OmpEndLoopDirective>>(
924               loopConstruct.t)) {
925     const auto &clauseList =
926         std::get<Fortran::parser::OmpClauseList>((*endClauseList).t);
927     for (const Fortran::parser::OmpClause &clause : clauseList.v)
928       if (std::get_if<Fortran::parser::OmpClause::Nowait>(&clause.u))
929         wsLoopOp.nowaitAttr(firOpBuilder.getUnitAttr());
930   }
931 
932   createBodyOfOp<omp::WsLoopOp>(wsLoopOp, converter, currentLocation, eval,
933                                 &loopOpClauseList, iv);
934 }
935 
936 static void
937 genOMP(Fortran::lower::AbstractConverter &converter,
938        Fortran::lower::pft::Evaluation &eval,
939        const Fortran::parser::OpenMPCriticalConstruct &criticalConstruct) {
940   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
941   mlir::Location currentLocation = converter.getCurrentLocation();
942   std::string name;
943   const Fortran::parser::OmpCriticalDirective &cd =
944       std::get<Fortran::parser::OmpCriticalDirective>(criticalConstruct.t);
945   if (std::get<std::optional<Fortran::parser::Name>>(cd.t).has_value()) {
946     name =
947         std::get<std::optional<Fortran::parser::Name>>(cd.t).value().ToString();
948   }
949 
950   uint64_t hint = 0;
951   const auto &clauseList = std::get<Fortran::parser::OmpClauseList>(cd.t);
952   for (const Fortran::parser::OmpClause &clause : clauseList.v)
953     if (auto hintClause =
954             std::get_if<Fortran::parser::OmpClause::Hint>(&clause.u)) {
955       const auto *expr = Fortran::semantics::GetExpr(hintClause->v);
956       hint = *Fortran::evaluate::ToInt64(*expr);
957       break;
958     }
959 
960   mlir::omp::CriticalOp criticalOp = [&]() {
961     if (name.empty()) {
962       return firOpBuilder.create<mlir::omp::CriticalOp>(currentLocation,
963                                                         FlatSymbolRefAttr());
964     } else {
965       mlir::ModuleOp module = firOpBuilder.getModule();
966       mlir::OpBuilder modBuilder(module.getBodyRegion());
967       auto global = module.lookupSymbol<mlir::omp::CriticalDeclareOp>(name);
968       if (!global)
969         global = modBuilder.create<mlir::omp::CriticalDeclareOp>(
970             currentLocation, name, hint);
971       return firOpBuilder.create<mlir::omp::CriticalOp>(
972           currentLocation, mlir::FlatSymbolRefAttr::get(
973                                firOpBuilder.getContext(), global.sym_name()));
974     }
975   }();
976   createBodyOfOp<omp::CriticalOp>(criticalOp, converter, currentLocation, eval);
977 }
978 
979 static void
980 genOMP(Fortran::lower::AbstractConverter &converter,
981        Fortran::lower::pft::Evaluation &eval,
982        const Fortran::parser::OpenMPSectionConstruct &sectionConstruct) {
983 
984   auto &firOpBuilder = converter.getFirOpBuilder();
985   auto currentLocation = converter.getCurrentLocation();
986   mlir::omp::SectionOp sectionOp =
987       firOpBuilder.create<mlir::omp::SectionOp>(currentLocation);
988   createBodyOfOp<omp::SectionOp>(sectionOp, converter, currentLocation, eval);
989 }
990 
991 // TODO: Add support for reduction
992 static void
993 genOMP(Fortran::lower::AbstractConverter &converter,
994        Fortran::lower::pft::Evaluation &eval,
995        const Fortran::parser::OpenMPSectionsConstruct &sectionsConstruct) {
996   auto &firOpBuilder = converter.getFirOpBuilder();
997   auto currentLocation = converter.getCurrentLocation();
998   SmallVector<Value> reductionVars, allocateOperands, allocatorOperands;
999   mlir::UnitAttr noWaitClauseOperand;
1000   const auto &sectionsClauseList = std::get<Fortran::parser::OmpClauseList>(
1001       std::get<Fortran::parser::OmpBeginSectionsDirective>(sectionsConstruct.t)
1002           .t);
1003   for (const Fortran::parser::OmpClause &clause : sectionsClauseList.v) {
1004 
1005     // Reduction Clause
1006     if (std::get_if<Fortran::parser::OmpClause::Reduction>(&clause.u)) {
1007       TODO(currentLocation, "OMPC_Reduction");
1008 
1009       // Allocate clause
1010     } else if (const auto &allocateClause =
1011                    std::get_if<Fortran::parser::OmpClause::Allocate>(
1012                        &clause.u)) {
1013       genAllocateClause(converter, allocateClause->v, allocatorOperands,
1014                         allocateOperands);
1015     }
1016   }
1017   const auto &endSectionsClauseList =
1018       std::get<Fortran::parser::OmpEndSectionsDirective>(sectionsConstruct.t);
1019   const auto &clauseList =
1020       std::get<Fortran::parser::OmpClauseList>(endSectionsClauseList.t);
1021   for (const auto &clause : clauseList.v) {
1022     // Nowait clause
1023     if (std::get_if<Fortran::parser::OmpClause::Nowait>(&clause.u)) {
1024       noWaitClauseOperand = firOpBuilder.getUnitAttr();
1025     }
1026   }
1027 
1028   llvm::omp::Directive dir =
1029       std::get<Fortran::parser::OmpSectionsDirective>(
1030           std::get<Fortran::parser::OmpBeginSectionsDirective>(
1031               sectionsConstruct.t)
1032               .t)
1033           .v;
1034 
1035   // Parallel Sections Construct
1036   if (dir == llvm::omp::Directive::OMPD_parallel_sections) {
1037     createCombinedParallelOp<Fortran::parser::OmpBeginSectionsDirective>(
1038         converter, eval,
1039         std::get<Fortran::parser::OmpBeginSectionsDirective>(
1040             sectionsConstruct.t));
1041     auto sectionsOp = firOpBuilder.create<mlir::omp::SectionsOp>(
1042         currentLocation, /*reduction_vars*/ ValueRange(),
1043         /*reductions=*/nullptr, allocateOperands, allocatorOperands,
1044         /*nowait=*/nullptr);
1045     createBodyOfOp(sectionsOp, converter, currentLocation, eval);
1046 
1047     // Sections Construct
1048   } else if (dir == llvm::omp::Directive::OMPD_sections) {
1049     auto sectionsOp = firOpBuilder.create<mlir::omp::SectionsOp>(
1050         currentLocation, reductionVars, /*reductions = */ nullptr,
1051         allocateOperands, allocatorOperands, noWaitClauseOperand);
1052     createBodyOfOp<omp::SectionsOp>(sectionsOp, converter, currentLocation,
1053                                     eval);
1054   }
1055 }
1056 
1057 static void genOmpAtomicHintAndMemoryOrderClauses(
1058     Fortran::lower::AbstractConverter &converter,
1059     const Fortran::parser::OmpAtomicClauseList &clauseList,
1060     mlir::IntegerAttr &hint,
1061     mlir::omp::ClauseMemoryOrderKindAttr &memory_order) {
1062   auto &firOpBuilder = converter.getFirOpBuilder();
1063   for (const auto &clause : clauseList.v) {
1064     if (auto ompClause = std::get_if<Fortran::parser::OmpClause>(&clause.u)) {
1065       if (auto hintClause =
1066               std::get_if<Fortran::parser::OmpClause::Hint>(&ompClause->u)) {
1067         const auto *expr = Fortran::semantics::GetExpr(hintClause->v);
1068         uint64_t hintExprValue = *Fortran::evaluate::ToInt64(*expr);
1069         hint = firOpBuilder.getI64IntegerAttr(hintExprValue);
1070       }
1071     } else if (auto ompMemoryOrderClause =
1072                    std::get_if<Fortran::parser::OmpMemoryOrderClause>(
1073                        &clause.u)) {
1074       if (std::get_if<Fortran::parser::OmpClause::Acquire>(
1075               &ompMemoryOrderClause->v.u)) {
1076         memory_order = mlir::omp::ClauseMemoryOrderKindAttr::get(
1077             firOpBuilder.getContext(), omp::ClauseMemoryOrderKind::Acquire);
1078       } else if (std::get_if<Fortran::parser::OmpClause::Relaxed>(
1079                      &ompMemoryOrderClause->v.u)) {
1080         memory_order = mlir::omp::ClauseMemoryOrderKindAttr::get(
1081             firOpBuilder.getContext(), omp::ClauseMemoryOrderKind::Relaxed);
1082       } else if (std::get_if<Fortran::parser::OmpClause::SeqCst>(
1083                      &ompMemoryOrderClause->v.u)) {
1084         memory_order = mlir::omp::ClauseMemoryOrderKindAttr::get(
1085             firOpBuilder.getContext(), omp::ClauseMemoryOrderKind::Seq_cst);
1086       } else if (std::get_if<Fortran::parser::OmpClause::Release>(
1087                      &ompMemoryOrderClause->v.u)) {
1088         memory_order = mlir::omp::ClauseMemoryOrderKindAttr::get(
1089             firOpBuilder.getContext(), omp::ClauseMemoryOrderKind::Release);
1090       }
1091     }
1092   }
1093 }
1094 
1095 static void
1096 genOmpAtomicWrite(Fortran::lower::AbstractConverter &converter,
1097                   Fortran::lower::pft::Evaluation &eval,
1098                   const Fortran::parser::OmpAtomicWrite &atomicWrite) {
1099   auto &firOpBuilder = converter.getFirOpBuilder();
1100   auto currentLocation = converter.getCurrentLocation();
1101   // Get the value and address of atomic write operands.
1102   const Fortran::parser::OmpAtomicClauseList &rightHandClauseList =
1103       std::get<2>(atomicWrite.t);
1104   const Fortran::parser::OmpAtomicClauseList &leftHandClauseList =
1105       std::get<0>(atomicWrite.t);
1106   const auto &assignmentStmtExpr =
1107       std::get<Fortran::parser::Expr>(std::get<3>(atomicWrite.t).statement.t);
1108   const auto &assignmentStmtVariable = std::get<Fortran::parser::Variable>(
1109       std::get<3>(atomicWrite.t).statement.t);
1110   Fortran::lower::StatementContext stmtCtx;
1111   mlir::Value value = fir::getBase(converter.genExprValue(
1112       *Fortran::semantics::GetExpr(assignmentStmtExpr), stmtCtx));
1113   mlir::Value address = fir::getBase(converter.genExprAddr(
1114       *Fortran::semantics::GetExpr(assignmentStmtVariable), stmtCtx));
1115   // If no hint clause is specified, the effect is as if
1116   // hint(omp_sync_hint_none) had been specified.
1117   mlir::IntegerAttr hint = nullptr;
1118   mlir::omp::ClauseMemoryOrderKindAttr memory_order = nullptr;
1119   genOmpAtomicHintAndMemoryOrderClauses(converter, leftHandClauseList, hint,
1120                                         memory_order);
1121   genOmpAtomicHintAndMemoryOrderClauses(converter, rightHandClauseList, hint,
1122                                         memory_order);
1123   firOpBuilder.create<mlir::omp::AtomicWriteOp>(currentLocation, address, value,
1124                                                 hint, memory_order);
1125 }
1126 
1127 static void genOmpAtomicRead(Fortran::lower::AbstractConverter &converter,
1128                              Fortran::lower::pft::Evaluation &eval,
1129                              const Fortran::parser::OmpAtomicRead &atomicRead) {
1130   auto &firOpBuilder = converter.getFirOpBuilder();
1131   auto currentLocation = converter.getCurrentLocation();
1132   // Get the address of atomic read operands.
1133   const Fortran::parser::OmpAtomicClauseList &rightHandClauseList =
1134       std::get<2>(atomicRead.t);
1135   const Fortran::parser::OmpAtomicClauseList &leftHandClauseList =
1136       std::get<0>(atomicRead.t);
1137   const auto &assignmentStmtExpr =
1138       std::get<Fortran::parser::Expr>(std::get<3>(atomicRead.t).statement.t);
1139   const auto &assignmentStmtVariable = std::get<Fortran::parser::Variable>(
1140       std::get<3>(atomicRead.t).statement.t);
1141   Fortran::lower::StatementContext stmtCtx;
1142   mlir::Value from_address = fir::getBase(converter.genExprAddr(
1143       *Fortran::semantics::GetExpr(assignmentStmtExpr), stmtCtx));
1144   mlir::Value to_address = fir::getBase(converter.genExprAddr(
1145       *Fortran::semantics::GetExpr(assignmentStmtVariable), stmtCtx));
1146   // If no hint clause is specified, the effect is as if
1147   // hint(omp_sync_hint_none) had been specified.
1148   mlir::IntegerAttr hint = nullptr;
1149   mlir::omp::ClauseMemoryOrderKindAttr memory_order = nullptr;
1150   genOmpAtomicHintAndMemoryOrderClauses(converter, leftHandClauseList, hint,
1151                                         memory_order);
1152   genOmpAtomicHintAndMemoryOrderClauses(converter, rightHandClauseList, hint,
1153                                         memory_order);
1154   firOpBuilder.create<mlir::omp::AtomicReadOp>(currentLocation, from_address,
1155                                                to_address, hint, memory_order);
1156 }
1157 
1158 static void
1159 genOMP(Fortran::lower::AbstractConverter &converter,
1160        Fortran::lower::pft::Evaluation &eval,
1161        const Fortran::parser::OpenMPAtomicConstruct &atomicConstruct) {
1162   std::visit(Fortran::common::visitors{
1163                  [&](const Fortran::parser::OmpAtomicRead &atomicRead) {
1164                    genOmpAtomicRead(converter, eval, atomicRead);
1165                  },
1166                  [&](const Fortran::parser::OmpAtomicWrite &atomicWrite) {
1167                    genOmpAtomicWrite(converter, eval, atomicWrite);
1168                  },
1169                  [&](const auto &) {
1170                    TODO(converter.getCurrentLocation(),
1171                         "Atomic update & capture");
1172                  },
1173              },
1174              atomicConstruct.u);
1175 }
1176 
1177 void Fortran::lower::genOpenMPConstruct(
1178     Fortran::lower::AbstractConverter &converter,
1179     Fortran::lower::pft::Evaluation &eval,
1180     const Fortran::parser::OpenMPConstruct &ompConstruct) {
1181 
1182   std::visit(
1183       common::visitors{
1184           [&](const Fortran::parser::OpenMPStandaloneConstruct
1185                   &standaloneConstruct) {
1186             genOMP(converter, eval, standaloneConstruct);
1187           },
1188           [&](const Fortran::parser::OpenMPSectionsConstruct
1189                   &sectionsConstruct) {
1190             genOMP(converter, eval, sectionsConstruct);
1191           },
1192           [&](const Fortran::parser::OpenMPSectionConstruct &sectionConstruct) {
1193             genOMP(converter, eval, sectionConstruct);
1194           },
1195           [&](const Fortran::parser::OpenMPLoopConstruct &loopConstruct) {
1196             genOMP(converter, eval, loopConstruct);
1197           },
1198           [&](const Fortran::parser::OpenMPDeclarativeAllocate
1199                   &execAllocConstruct) {
1200             TODO(converter.getCurrentLocation(), "OpenMPDeclarativeAllocate");
1201           },
1202           [&](const Fortran::parser::OpenMPExecutableAllocate
1203                   &execAllocConstruct) {
1204             TODO(converter.getCurrentLocation(), "OpenMPExecutableAllocate");
1205           },
1206           [&](const Fortran::parser::OpenMPBlockConstruct &blockConstruct) {
1207             genOMP(converter, eval, blockConstruct);
1208           },
1209           [&](const Fortran::parser::OpenMPAtomicConstruct &atomicConstruct) {
1210             genOMP(converter, eval, atomicConstruct);
1211           },
1212           [&](const Fortran::parser::OpenMPCriticalConstruct
1213                   &criticalConstruct) {
1214             genOMP(converter, eval, criticalConstruct);
1215           },
1216       },
1217       ompConstruct.u);
1218 }
1219 
1220 void Fortran::lower::genThreadprivateOp(
1221     Fortran::lower::AbstractConverter &converter,
1222     const Fortran::lower::pft::Variable &var) {
1223   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
1224   mlir::Location currentLocation = converter.getCurrentLocation();
1225 
1226   const Fortran::semantics::Symbol &sym = var.getSymbol();
1227   mlir::Value symThreadprivateValue;
1228   if (const Fortran::semantics::Symbol *common =
1229           Fortran::semantics::FindCommonBlockContaining(sym.GetUltimate())) {
1230     mlir::Value commonValue = converter.getSymbolAddress(*common);
1231     if (mlir::isa<mlir::omp::ThreadprivateOp>(commonValue.getDefiningOp())) {
1232       // Generate ThreadprivateOp for a common block instead of its members and
1233       // only do it once for a common block.
1234       return;
1235     }
1236     // Generate ThreadprivateOp and rebind the common block.
1237     mlir::Value commonThreadprivateValue =
1238         firOpBuilder.create<mlir::omp::ThreadprivateOp>(
1239             currentLocation, commonValue.getType(), commonValue);
1240     converter.bindSymbol(*common, commonThreadprivateValue);
1241     // Generate the threadprivate value for the common block member.
1242     symThreadprivateValue =
1243         genCommonBlockMember(converter, sym, commonThreadprivateValue);
1244   } else {
1245     mlir::Value symValue = converter.getSymbolAddress(sym);
1246     symThreadprivateValue = firOpBuilder.create<mlir::omp::ThreadprivateOp>(
1247         currentLocation, symValue.getType(), symValue);
1248   }
1249 
1250   fir::ExtendedValue sexv = converter.getSymbolExtendedValue(sym);
1251   fir::ExtendedValue symThreadprivateExv =
1252       getExtendedValue(sexv, symThreadprivateValue);
1253   converter.bindSymbol(sym, symThreadprivateExv);
1254 }
1255 
1256 void Fortran::lower::genOpenMPDeclarativeConstruct(
1257     Fortran::lower::AbstractConverter &converter,
1258     Fortran::lower::pft::Evaluation &eval,
1259     const Fortran::parser::OpenMPDeclarativeConstruct &ompDeclConstruct) {
1260 
1261   std::visit(
1262       common::visitors{
1263           [&](const Fortran::parser::OpenMPDeclarativeAllocate
1264                   &declarativeAllocate) {
1265             TODO(converter.getCurrentLocation(), "OpenMPDeclarativeAllocate");
1266           },
1267           [&](const Fortran::parser::OpenMPDeclareReductionConstruct
1268                   &declareReductionConstruct) {
1269             TODO(converter.getCurrentLocation(),
1270                  "OpenMPDeclareReductionConstruct");
1271           },
1272           [&](const Fortran::parser::OpenMPDeclareSimdConstruct
1273                   &declareSimdConstruct) {
1274             TODO(converter.getCurrentLocation(), "OpenMPDeclareSimdConstruct");
1275           },
1276           [&](const Fortran::parser::OpenMPDeclareTargetConstruct
1277                   &declareTargetConstruct) {
1278             TODO(converter.getCurrentLocation(),
1279                  "OpenMPDeclareTargetConstruct");
1280           },
1281           [&](const Fortran::parser::OpenMPThreadprivate &threadprivate) {
1282             // The directive is lowered when instantiating the variable to
1283             // support the case of threadprivate variable declared in module.
1284           },
1285       },
1286       ompDeclConstruct.u);
1287 }
1288