xref: /llvm-project/flang/lib/Lower/OpenMP/ReductionProcessor.cpp (revision 8cc34fadec71c358c2c69bbca236294afb259e02) 
1 //===-- ReductionProcessor.cpp ----------------------------------*- C++ -*-===//
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 "ReductionProcessor.h"
14 
15 #include "flang/Lower/AbstractConverter.h"
16 #include "flang/Lower/ConvertType.h"
17 #include "flang/Lower/SymbolMap.h"
18 #include "flang/Optimizer/Builder/Complex.h"
19 #include "flang/Optimizer/Builder/HLFIRTools.h"
20 #include "flang/Optimizer/Builder/Todo.h"
21 #include "flang/Optimizer/Dialect/FIRType.h"
22 #include "flang/Optimizer/HLFIR/HLFIROps.h"
23 #include "flang/Optimizer/Support/FatalError.h"
24 #include "flang/Parser/tools.h"
25 #include "mlir/Dialect/OpenMP/OpenMPDialect.h"
26 #include "llvm/Support/CommandLine.h"
27 
28 static llvm::cl::opt<bool> forceByrefReduction(
29     "force-byref-reduction",
30     llvm::cl::desc("Pass all reduction arguments by reference"),
31     llvm::cl::Hidden);
32 
33 namespace Fortran {
34 namespace lower {
35 namespace omp {
36 
37 ReductionProcessor::ReductionIdentifier ReductionProcessor::getReductionType(
38     const omp::clause::ProcedureDesignator &pd) {
39   auto redType = llvm::StringSwitch<std::optional<ReductionIdentifier>>(
40                      getRealName(pd.v.id()).ToString())
41                      .Case("max", ReductionIdentifier::MAX)
42                      .Case("min", ReductionIdentifier::MIN)
43                      .Case("iand", ReductionIdentifier::IAND)
44                      .Case("ior", ReductionIdentifier::IOR)
45                      .Case("ieor", ReductionIdentifier::IEOR)
46                      .Default(std::nullopt);
47   assert(redType && "Invalid Reduction");
48   return *redType;
49 }
50 
51 ReductionProcessor::ReductionIdentifier ReductionProcessor::getReductionType(
52     omp::clause::DefinedOperator::IntrinsicOperator intrinsicOp) {
53   switch (intrinsicOp) {
54   case omp::clause::DefinedOperator::IntrinsicOperator::Add:
55     return ReductionIdentifier::ADD;
56   case omp::clause::DefinedOperator::IntrinsicOperator::Subtract:
57     return ReductionIdentifier::SUBTRACT;
58   case omp::clause::DefinedOperator::IntrinsicOperator::Multiply:
59     return ReductionIdentifier::MULTIPLY;
60   case omp::clause::DefinedOperator::IntrinsicOperator::AND:
61     return ReductionIdentifier::AND;
62   case omp::clause::DefinedOperator::IntrinsicOperator::EQV:
63     return ReductionIdentifier::EQV;
64   case omp::clause::DefinedOperator::IntrinsicOperator::OR:
65     return ReductionIdentifier::OR;
66   case omp::clause::DefinedOperator::IntrinsicOperator::NEQV:
67     return ReductionIdentifier::NEQV;
68   default:
69     llvm_unreachable("unexpected intrinsic operator in reduction");
70   }
71 }
72 
73 bool ReductionProcessor::supportedIntrinsicProcReduction(
74     const omp::clause::ProcedureDesignator &pd) {
75   Fortran::semantics::Symbol *sym = pd.v.id();
76   if (!sym->GetUltimate().attrs().test(Fortran::semantics::Attr::INTRINSIC))
77     return false;
78   auto redType = llvm::StringSwitch<bool>(getRealName(sym).ToString())
79                      .Case("max", true)
80                      .Case("min", true)
81                      .Case("iand", true)
82                      .Case("ior", true)
83                      .Case("ieor", true)
84                      .Default(false);
85   return redType;
86 }
87 
88 std::string
89 ReductionProcessor::getReductionName(llvm::StringRef name,
90                                      const fir::KindMapping &kindMap,
91                                      mlir::Type ty, bool isByRef) {
92   ty = fir::unwrapRefType(ty);
93 
94   // extra string to distinguish reduction functions for variables passed by
95   // reference
96   llvm::StringRef byrefAddition{""};
97   if (isByRef)
98     byrefAddition = "_byref";
99 
100   return fir::getTypeAsString(ty, kindMap, (name + byrefAddition).str());
101 }
102 
103 std::string ReductionProcessor::getReductionName(
104     omp::clause::DefinedOperator::IntrinsicOperator intrinsicOp,
105     const fir::KindMapping &kindMap, mlir::Type ty, bool isByRef) {
106   std::string reductionName;
107 
108   switch (intrinsicOp) {
109   case omp::clause::DefinedOperator::IntrinsicOperator::Add:
110     reductionName = "add_reduction";
111     break;
112   case omp::clause::DefinedOperator::IntrinsicOperator::Multiply:
113     reductionName = "multiply_reduction";
114     break;
115   case omp::clause::DefinedOperator::IntrinsicOperator::AND:
116     return "and_reduction";
117   case omp::clause::DefinedOperator::IntrinsicOperator::EQV:
118     return "eqv_reduction";
119   case omp::clause::DefinedOperator::IntrinsicOperator::OR:
120     return "or_reduction";
121   case omp::clause::DefinedOperator::IntrinsicOperator::NEQV:
122     return "neqv_reduction";
123   default:
124     reductionName = "other_reduction";
125     break;
126   }
127 
128   return getReductionName(reductionName, kindMap, ty, isByRef);
129 }
130 
131 mlir::Value
132 ReductionProcessor::getReductionInitValue(mlir::Location loc, mlir::Type type,
133                                           ReductionIdentifier redId,
134                                           fir::FirOpBuilder &builder) {
135   type = fir::unwrapRefType(type);
136   if (!fir::isa_integer(type) && !fir::isa_real(type) &&
137       !fir::isa_complex(type) && !mlir::isa<fir::LogicalType>(type))
138     TODO(loc, "Reduction of some types is not supported");
139   switch (redId) {
140   case ReductionIdentifier::MAX: {
141     if (auto ty = type.dyn_cast<mlir::FloatType>()) {
142       const llvm::fltSemantics &sem = ty.getFloatSemantics();
143       return builder.createRealConstant(
144           loc, type, llvm::APFloat::getLargest(sem, /*Negative=*/true));
145     }
146     unsigned bits = type.getIntOrFloatBitWidth();
147     int64_t minInt = llvm::APInt::getSignedMinValue(bits).getSExtValue();
148     return builder.createIntegerConstant(loc, type, minInt);
149   }
150   case ReductionIdentifier::MIN: {
151     if (auto ty = type.dyn_cast<mlir::FloatType>()) {
152       const llvm::fltSemantics &sem = ty.getFloatSemantics();
153       return builder.createRealConstant(
154           loc, type, llvm::APFloat::getLargest(sem, /*Negative=*/false));
155     }
156     unsigned bits = type.getIntOrFloatBitWidth();
157     int64_t maxInt = llvm::APInt::getSignedMaxValue(bits).getSExtValue();
158     return builder.createIntegerConstant(loc, type, maxInt);
159   }
160   case ReductionIdentifier::IOR: {
161     unsigned bits = type.getIntOrFloatBitWidth();
162     int64_t zeroInt = llvm::APInt::getZero(bits).getSExtValue();
163     return builder.createIntegerConstant(loc, type, zeroInt);
164   }
165   case ReductionIdentifier::IEOR: {
166     unsigned bits = type.getIntOrFloatBitWidth();
167     int64_t zeroInt = llvm::APInt::getZero(bits).getSExtValue();
168     return builder.createIntegerConstant(loc, type, zeroInt);
169   }
170   case ReductionIdentifier::IAND: {
171     unsigned bits = type.getIntOrFloatBitWidth();
172     int64_t allOnInt = llvm::APInt::getAllOnes(bits).getSExtValue();
173     return builder.createIntegerConstant(loc, type, allOnInt);
174   }
175   case ReductionIdentifier::ADD:
176   case ReductionIdentifier::MULTIPLY:
177   case ReductionIdentifier::AND:
178   case ReductionIdentifier::OR:
179   case ReductionIdentifier::EQV:
180   case ReductionIdentifier::NEQV:
181     if (auto cplxTy = mlir::dyn_cast<fir::ComplexType>(type)) {
182       mlir::Type realTy =
183           Fortran::lower::convertReal(builder.getContext(), cplxTy.getFKind());
184       mlir::Value initRe = builder.createRealConstant(
185           loc, realTy, getOperationIdentity(redId, loc));
186       mlir::Value initIm = builder.createRealConstant(loc, realTy, 0);
187 
188       return fir::factory::Complex{builder, loc}.createComplex(type, initRe,
189                                                                initIm);
190     }
191     if (type.isa<mlir::FloatType>())
192       return builder.create<mlir::arith::ConstantOp>(
193           loc, type,
194           builder.getFloatAttr(type, (double)getOperationIdentity(redId, loc)));
195 
196     if (type.isa<fir::LogicalType>()) {
197       mlir::Value intConst = builder.create<mlir::arith::ConstantOp>(
198           loc, builder.getI1Type(),
199           builder.getIntegerAttr(builder.getI1Type(),
200                                  getOperationIdentity(redId, loc)));
201       return builder.createConvert(loc, type, intConst);
202     }
203 
204     return builder.create<mlir::arith::ConstantOp>(
205         loc, type,
206         builder.getIntegerAttr(type, getOperationIdentity(redId, loc)));
207   case ReductionIdentifier::ID:
208   case ReductionIdentifier::USER_DEF_OP:
209   case ReductionIdentifier::SUBTRACT:
210     TODO(loc, "Reduction of some identifier types is not supported");
211   }
212   llvm_unreachable("Unhandled Reduction identifier : getReductionInitValue");
213 }
214 
215 mlir::Value ReductionProcessor::createScalarCombiner(
216     fir::FirOpBuilder &builder, mlir::Location loc, ReductionIdentifier redId,
217     mlir::Type type, mlir::Value op1, mlir::Value op2) {
218   mlir::Value reductionOp;
219   type = fir::unwrapRefType(type);
220   switch (redId) {
221   case ReductionIdentifier::MAX:
222     reductionOp =
223         getReductionOperation<mlir::arith::MaxNumFOp, mlir::arith::MaxSIOp>(
224             builder, type, loc, op1, op2);
225     break;
226   case ReductionIdentifier::MIN:
227     reductionOp =
228         getReductionOperation<mlir::arith::MinNumFOp, mlir::arith::MinSIOp>(
229             builder, type, loc, op1, op2);
230     break;
231   case ReductionIdentifier::IOR:
232     assert((type.isIntOrIndex()) && "only integer is expected");
233     reductionOp = builder.create<mlir::arith::OrIOp>(loc, op1, op2);
234     break;
235   case ReductionIdentifier::IEOR:
236     assert((type.isIntOrIndex()) && "only integer is expected");
237     reductionOp = builder.create<mlir::arith::XOrIOp>(loc, op1, op2);
238     break;
239   case ReductionIdentifier::IAND:
240     assert((type.isIntOrIndex()) && "only integer is expected");
241     reductionOp = builder.create<mlir::arith::AndIOp>(loc, op1, op2);
242     break;
243   case ReductionIdentifier::ADD:
244     reductionOp =
245         getReductionOperation<mlir::arith::AddFOp, mlir::arith::AddIOp,
246                               fir::AddcOp>(builder, type, loc, op1, op2);
247     break;
248   case ReductionIdentifier::MULTIPLY:
249     reductionOp =
250         getReductionOperation<mlir::arith::MulFOp, mlir::arith::MulIOp,
251                               fir::MulcOp>(builder, type, loc, op1, op2);
252     break;
253   case ReductionIdentifier::AND: {
254     mlir::Value op1I1 = builder.createConvert(loc, builder.getI1Type(), op1);
255     mlir::Value op2I1 = builder.createConvert(loc, builder.getI1Type(), op2);
256 
257     mlir::Value andiOp = builder.create<mlir::arith::AndIOp>(loc, op1I1, op2I1);
258 
259     reductionOp = builder.createConvert(loc, type, andiOp);
260     break;
261   }
262   case ReductionIdentifier::OR: {
263     mlir::Value op1I1 = builder.createConvert(loc, builder.getI1Type(), op1);
264     mlir::Value op2I1 = builder.createConvert(loc, builder.getI1Type(), op2);
265 
266     mlir::Value oriOp = builder.create<mlir::arith::OrIOp>(loc, op1I1, op2I1);
267 
268     reductionOp = builder.createConvert(loc, type, oriOp);
269     break;
270   }
271   case ReductionIdentifier::EQV: {
272     mlir::Value op1I1 = builder.createConvert(loc, builder.getI1Type(), op1);
273     mlir::Value op2I1 = builder.createConvert(loc, builder.getI1Type(), op2);
274 
275     mlir::Value cmpiOp = builder.create<mlir::arith::CmpIOp>(
276         loc, mlir::arith::CmpIPredicate::eq, op1I1, op2I1);
277 
278     reductionOp = builder.createConvert(loc, type, cmpiOp);
279     break;
280   }
281   case ReductionIdentifier::NEQV: {
282     mlir::Value op1I1 = builder.createConvert(loc, builder.getI1Type(), op1);
283     mlir::Value op2I1 = builder.createConvert(loc, builder.getI1Type(), op2);
284 
285     mlir::Value cmpiOp = builder.create<mlir::arith::CmpIOp>(
286         loc, mlir::arith::CmpIPredicate::ne, op1I1, op2I1);
287 
288     reductionOp = builder.createConvert(loc, type, cmpiOp);
289     break;
290   }
291   default:
292     TODO(loc, "Reduction of some intrinsic operators is not supported");
293   }
294 
295   return reductionOp;
296 }
297 
298 /// Create reduction combiner region for reduction variables which are boxed
299 /// arrays
300 static void genBoxCombiner(fir::FirOpBuilder &builder, mlir::Location loc,
301                            ReductionProcessor::ReductionIdentifier redId,
302                            fir::BaseBoxType boxTy, mlir::Value lhs,
303                            mlir::Value rhs) {
304   fir::SequenceType seqTy = mlir::dyn_cast_or_null<fir::SequenceType>(
305       fir::unwrapRefType(boxTy.getEleTy()));
306   fir::HeapType heapTy =
307       mlir::dyn_cast_or_null<fir::HeapType>(boxTy.getEleTy());
308   if ((!seqTy || seqTy.hasUnknownShape()) && !heapTy)
309     TODO(loc, "Unsupported boxed type in OpenMP reduction");
310 
311   // load fir.ref<fir.box<...>>
312   mlir::Value lhsAddr = lhs;
313   lhs = builder.create<fir::LoadOp>(loc, lhs);
314   rhs = builder.create<fir::LoadOp>(loc, rhs);
315 
316   if (heapTy && !seqTy) {
317     // get box contents (heap pointers)
318     lhs = builder.create<fir::BoxAddrOp>(loc, lhs);
319     rhs = builder.create<fir::BoxAddrOp>(loc, rhs);
320     mlir::Value lhsValAddr = lhs;
321 
322     // load heap pointers
323     lhs = builder.create<fir::LoadOp>(loc, lhs);
324     rhs = builder.create<fir::LoadOp>(loc, rhs);
325 
326     mlir::Value result = ReductionProcessor::createScalarCombiner(
327         builder, loc, redId, heapTy.getEleTy(), lhs, rhs);
328     builder.create<fir::StoreOp>(loc, result, lhsValAddr);
329     builder.create<mlir::omp::YieldOp>(loc, lhsAddr);
330     return;
331   }
332 
333   const unsigned rank = seqTy.getDimension();
334   llvm::SmallVector<mlir::Value> extents;
335   extents.reserve(rank);
336   llvm::SmallVector<mlir::Value> lbAndExtents;
337   lbAndExtents.reserve(rank * 2);
338 
339   // Get box lowerbounds and extents:
340   mlir::Type idxTy = builder.getIndexType();
341   for (unsigned i = 0; i < rank; ++i) {
342     // TODO: ideally we want to hoist box reads out of the critical section.
343     // We could do this by having box dimensions in block arguments like
344     // OpenACC does
345     mlir::Value dim = builder.createIntegerConstant(loc, idxTy, i);
346     auto dimInfo =
347         builder.create<fir::BoxDimsOp>(loc, idxTy, idxTy, idxTy, lhs, dim);
348     extents.push_back(dimInfo.getExtent());
349     lbAndExtents.push_back(dimInfo.getLowerBound());
350     lbAndExtents.push_back(dimInfo.getExtent());
351   }
352 
353   auto shapeShiftTy = fir::ShapeShiftType::get(builder.getContext(), rank);
354   auto shapeShift =
355       builder.create<fir::ShapeShiftOp>(loc, shapeShiftTy, lbAndExtents);
356 
357   // Iterate over array elements, applying the equivalent scalar reduction:
358 
359   // F2018 5.4.10.2: Unallocated allocatable variables may not be referenced
360   // and so no null check is needed here before indexing into the (possibly
361   // allocatable) arrays.
362 
363   // A hlfir::elemental here gets inlined with a temporary so create the
364   // loop nest directly.
365   // This function already controls all of the code in this region so we
366   // know this won't miss any opportuinties for clever elemental inlining
367   hlfir::LoopNest nest =
368       hlfir::genLoopNest(loc, builder, extents, /*isUnordered=*/true);
369   builder.setInsertionPointToStart(nest.innerLoop.getBody());
370   mlir::Type refTy = fir::ReferenceType::get(seqTy.getEleTy());
371   auto lhsEleAddr = builder.create<fir::ArrayCoorOp>(
372       loc, refTy, lhs, shapeShift, /*slice=*/mlir::Value{},
373       nest.oneBasedIndices, /*typeparms=*/mlir::ValueRange{});
374   auto rhsEleAddr = builder.create<fir::ArrayCoorOp>(
375       loc, refTy, rhs, shapeShift, /*slice=*/mlir::Value{},
376       nest.oneBasedIndices, /*typeparms=*/mlir::ValueRange{});
377   auto lhsEle = builder.create<fir::LoadOp>(loc, lhsEleAddr);
378   auto rhsEle = builder.create<fir::LoadOp>(loc, rhsEleAddr);
379   mlir::Value scalarReduction = ReductionProcessor::createScalarCombiner(
380       builder, loc, redId, refTy, lhsEle, rhsEle);
381   builder.create<fir::StoreOp>(loc, scalarReduction, lhsEleAddr);
382 
383   builder.setInsertionPointAfter(nest.outerLoop);
384   builder.create<mlir::omp::YieldOp>(loc, lhsAddr);
385 }
386 
387 // generate combiner region for reduction operations
388 static void genCombiner(fir::FirOpBuilder &builder, mlir::Location loc,
389                         ReductionProcessor::ReductionIdentifier redId,
390                         mlir::Type ty, mlir::Value lhs, mlir::Value rhs,
391                         bool isByRef) {
392   ty = fir::unwrapRefType(ty);
393 
394   if (fir::isa_trivial(ty)) {
395     mlir::Value lhsLoaded = builder.loadIfRef(loc, lhs);
396     mlir::Value rhsLoaded = builder.loadIfRef(loc, rhs);
397 
398     mlir::Value result = ReductionProcessor::createScalarCombiner(
399         builder, loc, redId, ty, lhsLoaded, rhsLoaded);
400     if (isByRef) {
401       builder.create<fir::StoreOp>(loc, result, lhs);
402       builder.create<mlir::omp::YieldOp>(loc, lhs);
403     } else {
404       builder.create<mlir::omp::YieldOp>(loc, result);
405     }
406     return;
407   }
408   // all arrays should have been boxed
409   if (auto boxTy = mlir::dyn_cast<fir::BaseBoxType>(ty)) {
410     genBoxCombiner(builder, loc, redId, boxTy, lhs, rhs);
411     return;
412   }
413 
414   TODO(loc, "OpenMP genCombiner for unsupported reduction variable type");
415 }
416 
417 static void
418 createReductionCleanupRegion(fir::FirOpBuilder &builder, mlir::Location loc,
419                              mlir::omp::DeclareReductionOp &reductionDecl) {
420   mlir::Type redTy = reductionDecl.getType();
421 
422   mlir::Region &cleanupRegion = reductionDecl.getCleanupRegion();
423   assert(cleanupRegion.empty());
424   mlir::Block *block =
425       builder.createBlock(&cleanupRegion, cleanupRegion.end(), {redTy}, {loc});
426   builder.setInsertionPointToEnd(block);
427 
428   auto typeError = [loc]() {
429     fir::emitFatalError(loc,
430                         "Attempt to create an omp reduction cleanup region "
431                         "for a type that wasn't allocated",
432                         /*genCrashDiag=*/true);
433   };
434 
435   mlir::Type valTy = fir::unwrapRefType(redTy);
436   if (auto boxTy = mlir::dyn_cast_or_null<fir::BaseBoxType>(valTy)) {
437     if (!mlir::isa<fir::HeapType>(boxTy.getEleTy())) {
438       mlir::Type innerTy = fir::extractSequenceType(boxTy);
439       if (!mlir::isa<fir::SequenceType>(innerTy))
440         typeError();
441     }
442 
443     mlir::Value arg = block->getArgument(0);
444     arg = builder.loadIfRef(loc, arg);
445     assert(mlir::isa<fir::BaseBoxType>(arg.getType()));
446 
447     // Deallocate box
448     // The FIR type system doesn't nesecarrily know that this is a mutable box
449     // if we allocated the thread local array on the heap to avoid looped stack
450     // allocations.
451     mlir::Value addr =
452         hlfir::genVariableRawAddress(loc, builder, hlfir::Entity{arg});
453     mlir::Value isAllocated = builder.genIsNotNullAddr(loc, addr);
454     fir::IfOp ifOp =
455         builder.create<fir::IfOp>(loc, isAllocated, /*withElseRegion=*/false);
456     builder.setInsertionPointToStart(&ifOp.getThenRegion().front());
457 
458     mlir::Value cast = builder.createConvert(
459         loc, fir::HeapType::get(fir::dyn_cast_ptrEleTy(addr.getType())), addr);
460     builder.create<fir::FreeMemOp>(loc, cast);
461 
462     builder.setInsertionPointAfter(ifOp);
463     builder.create<mlir::omp::YieldOp>(loc);
464     return;
465   }
466 
467   typeError();
468 }
469 
470 // like fir::unwrapSeqOrBoxedSeqType except it also works for non-sequence boxes
471 static mlir::Type unwrapSeqOrBoxedType(mlir::Type ty) {
472   if (auto seqTy = ty.dyn_cast<fir::SequenceType>())
473     return seqTy.getEleTy();
474   if (auto boxTy = ty.dyn_cast<fir::BaseBoxType>()) {
475     auto eleTy = fir::unwrapRefType(boxTy.getEleTy());
476     if (auto seqTy = eleTy.dyn_cast<fir::SequenceType>())
477       return seqTy.getEleTy();
478     return eleTy;
479   }
480   return ty;
481 }
482 
483 static mlir::Value
484 createReductionInitRegion(fir::FirOpBuilder &builder, mlir::Location loc,
485                           mlir::omp::DeclareReductionOp &reductionDecl,
486                           const ReductionProcessor::ReductionIdentifier redId,
487                           mlir::Type type, bool isByRef) {
488   mlir::Type ty = fir::unwrapRefType(type);
489   mlir::Value initValue = ReductionProcessor::getReductionInitValue(
490       loc, unwrapSeqOrBoxedType(ty), redId, builder);
491 
492   if (fir::isa_trivial(ty)) {
493     if (isByRef) {
494       mlir::Value alloca = builder.create<fir::AllocaOp>(loc, ty);
495       builder.createStoreWithConvert(loc, initValue, alloca);
496       return alloca;
497     }
498     // by val
499     return initValue;
500   }
501 
502   // check if an allocatable box is unallocated. If so, initialize the boxAlloca
503   // to be unallocated e.g.
504   // %box_alloca = fir.alloca !fir.box<!fir.heap<...>>
505   // %addr = fir.box_addr %box
506   // if (%addr == 0) {
507   //   %nullbox = fir.embox %addr
508   //   fir.store %nullbox to %box_alloca
509   // } else {
510   //   // ...
511   //   fir.store %something to %box_alloca
512   // }
513   // omp.yield %box_alloca
514   mlir::Value blockArg =
515       builder.loadIfRef(loc, builder.getBlock()->getArgument(0));
516   auto handleNullAllocatable = [&](mlir::Value boxAlloca) -> fir::IfOp {
517     mlir::Value addr = builder.create<fir::BoxAddrOp>(loc, blockArg);
518     mlir::Value isNotAllocated = builder.genIsNullAddr(loc, addr);
519     fir::IfOp ifOp = builder.create<fir::IfOp>(loc, isNotAllocated,
520                                                /*withElseRegion=*/true);
521     builder.setInsertionPointToStart(&ifOp.getThenRegion().front());
522     // just embox the null address and return
523     mlir::Value nullBox = builder.create<fir::EmboxOp>(loc, ty, addr);
524     builder.create<fir::StoreOp>(loc, nullBox, boxAlloca);
525     return ifOp;
526   };
527 
528   // all arrays are boxed
529   if (auto boxTy = mlir::dyn_cast_or_null<fir::BaseBoxType>(ty)) {
530     assert(isByRef && "passing boxes by value is unsupported");
531     bool isAllocatable = mlir::isa<fir::HeapType>(boxTy.getEleTy());
532     mlir::Value boxAlloca = builder.create<fir::AllocaOp>(loc, ty);
533     mlir::Type innerTy = fir::unwrapRefType(boxTy.getEleTy());
534     if (fir::isa_trivial(innerTy)) {
535       // boxed non-sequence value e.g. !fir.box<!fir.heap<i32>>
536       if (!isAllocatable)
537         TODO(loc, "Reduction of non-allocatable trivial typed box");
538 
539       fir::IfOp ifUnallocated = handleNullAllocatable(boxAlloca);
540 
541       builder.setInsertionPointToStart(&ifUnallocated.getElseRegion().front());
542       mlir::Value valAlloc = builder.create<fir::AllocMemOp>(loc, innerTy);
543       builder.createStoreWithConvert(loc, initValue, valAlloc);
544       mlir::Value box = builder.create<fir::EmboxOp>(loc, ty, valAlloc);
545       builder.create<fir::StoreOp>(loc, box, boxAlloca);
546 
547       auto insPt = builder.saveInsertionPoint();
548       createReductionCleanupRegion(builder, loc, reductionDecl);
549       builder.restoreInsertionPoint(insPt);
550       builder.setInsertionPointAfter(ifUnallocated);
551       return boxAlloca;
552     }
553     innerTy = fir::extractSequenceType(boxTy);
554     if (!mlir::isa<fir::SequenceType>(innerTy))
555       TODO(loc, "Unsupported boxed type for reduction");
556 
557     fir::IfOp ifUnallocated{nullptr};
558     if (isAllocatable) {
559       ifUnallocated = handleNullAllocatable(boxAlloca);
560       builder.setInsertionPointToStart(&ifUnallocated.getElseRegion().front());
561     }
562 
563     // Create the private copy from the initial fir.box:
564     hlfir::Entity source = hlfir::Entity{blockArg};
565 
566     // Allocating on the heap in case the whole reduction is nested inside of a
567     // loop
568     // TODO: compare performance here to using allocas - this could be made to
569     // work by inserting stacksave/stackrestore around the reduction in
570     // openmpirbuilder
571     auto [temp, needsDealloc] = createTempFromMold(loc, builder, source);
572     // if needsDealloc isn't statically false, add cleanup region. Always
573     // do this for allocatable boxes because they might have been re-allocated
574     // in the body of the loop/parallel region
575 
576     std::optional<int64_t> cstNeedsDealloc =
577         fir::getIntIfConstant(needsDealloc);
578     assert(cstNeedsDealloc.has_value() &&
579            "createTempFromMold decides this statically");
580     if (cstNeedsDealloc.has_value() && *cstNeedsDealloc != false) {
581       mlir::OpBuilder::InsertionGuard guard(builder);
582       createReductionCleanupRegion(builder, loc, reductionDecl);
583     } else {
584       assert(!isAllocatable && "Allocatable arrays must be heap allocated");
585     }
586 
587     // Put the temporary inside of a box:
588     hlfir::Entity box = hlfir::genVariableBox(loc, builder, temp);
589     // hlfir::genVariableBox removes fir.heap<> around the element type
590     mlir::Value convertedBox = builder.createConvert(loc, ty, box.getBase());
591     builder.create<hlfir::AssignOp>(loc, initValue, convertedBox);
592     builder.create<fir::StoreOp>(loc, convertedBox, boxAlloca);
593     if (ifUnallocated)
594       builder.setInsertionPointAfter(ifUnallocated);
595     return boxAlloca;
596   }
597 
598   TODO(loc, "createReductionInitRegion for unsupported type");
599 }
600 
601 mlir::omp::DeclareReductionOp ReductionProcessor::createDeclareReduction(
602     fir::FirOpBuilder &builder, llvm::StringRef reductionOpName,
603     const ReductionIdentifier redId, mlir::Type type, mlir::Location loc,
604     bool isByRef) {
605   mlir::OpBuilder::InsertionGuard guard(builder);
606   mlir::ModuleOp module = builder.getModule();
607 
608   assert(!reductionOpName.empty());
609 
610   auto decl =
611       module.lookupSymbol<mlir::omp::DeclareReductionOp>(reductionOpName);
612   if (decl)
613     return decl;
614 
615   mlir::OpBuilder modBuilder(module.getBodyRegion());
616   mlir::Type valTy = fir::unwrapRefType(type);
617   if (!isByRef)
618     type = valTy;
619 
620   decl = modBuilder.create<mlir::omp::DeclareReductionOp>(loc, reductionOpName,
621                                                           type);
622   builder.createBlock(&decl.getInitializerRegion(),
623                       decl.getInitializerRegion().end(), {type}, {loc});
624   builder.setInsertionPointToEnd(&decl.getInitializerRegion().back());
625 
626   mlir::Value init =
627       createReductionInitRegion(builder, loc, decl, redId, type, isByRef);
628   builder.create<mlir::omp::YieldOp>(loc, init);
629 
630   builder.createBlock(&decl.getReductionRegion(),
631                       decl.getReductionRegion().end(), {type, type},
632                       {loc, loc});
633 
634   builder.setInsertionPointToEnd(&decl.getReductionRegion().back());
635   mlir::Value op1 = decl.getReductionRegion().front().getArgument(0);
636   mlir::Value op2 = decl.getReductionRegion().front().getArgument(1);
637   genCombiner(builder, loc, redId, type, op1, op2, isByRef);
638 
639   return decl;
640 }
641 
642 // TODO: By-ref vs by-val reductions are currently toggled for the whole
643 //       operation (possibly effecting multiple reduction variables).
644 //       This could cause a problem with openmp target reductions because
645 //       by-ref trivial types may not be supported.
646 bool ReductionProcessor::doReductionByRef(
647     const llvm::SmallVectorImpl<mlir::Value> &reductionVars) {
648   if (reductionVars.empty())
649     return false;
650   if (forceByrefReduction)
651     return true;
652 
653   for (mlir::Value reductionVar : reductionVars) {
654     if (auto declare =
655             mlir::dyn_cast<hlfir::DeclareOp>(reductionVar.getDefiningOp()))
656       reductionVar = declare.getMemref();
657 
658     if (!fir::isa_trivial(fir::unwrapRefType(reductionVar.getType())))
659       return true;
660   }
661   return false;
662 }
663 
664 void ReductionProcessor::addDeclareReduction(
665     mlir::Location currentLocation,
666     Fortran::lower::AbstractConverter &converter,
667     const omp::clause::Reduction &reduction,
668     llvm::SmallVectorImpl<mlir::Value> &reductionVars,
669     llvm::SmallVectorImpl<mlir::Attribute> &reductionDeclSymbols,
670     llvm::SmallVectorImpl<const Fortran::semantics::Symbol *>
671         *reductionSymbols) {
672   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
673 
674   if (std::get<std::optional<omp::clause::Reduction::ReductionModifier>>(
675           reduction.t))
676     TODO(currentLocation, "Reduction modifiers are not supported");
677 
678   mlir::omp::DeclareReductionOp decl;
679   const auto &redOperatorList{
680       std::get<omp::clause::Reduction::ReductionIdentifiers>(reduction.t)};
681   assert(redOperatorList.size() == 1 && "Expecting single operator");
682   const auto &redOperator = redOperatorList.front();
683   const auto &objectList{std::get<omp::ObjectList>(reduction.t)};
684 
685   if (!std::holds_alternative<omp::clause::DefinedOperator>(redOperator.u)) {
686     if (const auto *reductionIntrinsic =
687             std::get_if<omp::clause::ProcedureDesignator>(&redOperator.u)) {
688       if (!ReductionProcessor::supportedIntrinsicProcReduction(
689               *reductionIntrinsic)) {
690         return;
691       }
692     } else {
693       return;
694     }
695   }
696 
697   // initial pass to collect all reduction vars so we can figure out if this
698   // should happen byref
699   fir::FirOpBuilder &builder = converter.getFirOpBuilder();
700   for (const Object &object : objectList) {
701     const Fortran::semantics::Symbol *symbol = object.id();
702     if (reductionSymbols)
703       reductionSymbols->push_back(symbol);
704     mlir::Value symVal = converter.getSymbolAddress(*symbol);
705     mlir::Type eleType;
706     auto refType = mlir::dyn_cast_or_null<fir::ReferenceType>(symVal.getType());
707     if (refType)
708       eleType = refType.getEleTy();
709     else
710       eleType = symVal.getType();
711 
712     // all arrays must be boxed so that we have convenient access to all the
713     // information needed to iterate over the array
714     if (mlir::isa<fir::SequenceType>(eleType)) {
715       // For Host associated symbols, use `SymbolBox` instead
716       Fortran::lower::SymbolBox symBox =
717           converter.lookupOneLevelUpSymbol(*symbol);
718       hlfir::Entity entity{symBox.getAddr()};
719       entity = genVariableBox(currentLocation, builder, entity);
720       mlir::Value box = entity.getBase();
721 
722       // Always pass the box by reference so that the OpenMP dialect
723       // verifiers don't need to know anything about fir.box
724       auto alloca =
725           builder.create<fir::AllocaOp>(currentLocation, box.getType());
726       builder.create<fir::StoreOp>(currentLocation, box, alloca);
727 
728       symVal = alloca;
729     } else if (mlir::isa<fir::BaseBoxType>(symVal.getType())) {
730       // boxed arrays are passed as values not by reference. Unfortunately,
731       // we can't pass a box by value to omp.redution_declare, so turn it
732       // into a reference
733 
734       auto alloca =
735           builder.create<fir::AllocaOp>(currentLocation, symVal.getType());
736       builder.create<fir::StoreOp>(currentLocation, symVal, alloca);
737       symVal = alloca;
738     } else if (auto declOp = symVal.getDefiningOp<hlfir::DeclareOp>()) {
739       symVal = declOp.getBase();
740     }
741 
742     // this isn't the same as the by-val and by-ref passing later in the
743     // pipeline. Both styles assume that the variable is a reference at
744     // this point
745     assert(mlir::isa<fir::ReferenceType>(symVal.getType()) &&
746            "reduction input var is a reference");
747 
748     reductionVars.push_back(symVal);
749   }
750   const bool isByRef = doReductionByRef(reductionVars);
751 
752   if (const auto &redDefinedOp =
753           std::get_if<omp::clause::DefinedOperator>(&redOperator.u)) {
754     const auto &intrinsicOp{
755         std::get<omp::clause::DefinedOperator::IntrinsicOperator>(
756             redDefinedOp->u)};
757     ReductionIdentifier redId = getReductionType(intrinsicOp);
758     switch (redId) {
759     case ReductionIdentifier::ADD:
760     case ReductionIdentifier::MULTIPLY:
761     case ReductionIdentifier::AND:
762     case ReductionIdentifier::EQV:
763     case ReductionIdentifier::OR:
764     case ReductionIdentifier::NEQV:
765       break;
766     default:
767       TODO(currentLocation,
768            "Reduction of some intrinsic operators is not supported");
769       break;
770     }
771 
772     for (mlir::Value symVal : reductionVars) {
773       auto redType = mlir::cast<fir::ReferenceType>(symVal.getType());
774       const auto &kindMap = firOpBuilder.getKindMap();
775       if (redType.getEleTy().isa<fir::LogicalType>())
776         decl = createDeclareReduction(firOpBuilder,
777                                       getReductionName(intrinsicOp, kindMap,
778                                                        firOpBuilder.getI1Type(),
779                                                        isByRef),
780                                       redId, redType, currentLocation, isByRef);
781       else
782         decl = createDeclareReduction(
783             firOpBuilder,
784             getReductionName(intrinsicOp, kindMap, redType, isByRef), redId,
785             redType, currentLocation, isByRef);
786       reductionDeclSymbols.push_back(mlir::SymbolRefAttr::get(
787           firOpBuilder.getContext(), decl.getSymName()));
788     }
789   } else if (const auto *reductionIntrinsic =
790                  std::get_if<omp::clause::ProcedureDesignator>(
791                      &redOperator.u)) {
792     if (ReductionProcessor::supportedIntrinsicProcReduction(
793             *reductionIntrinsic)) {
794       ReductionProcessor::ReductionIdentifier redId =
795           ReductionProcessor::getReductionType(*reductionIntrinsic);
796       for (const Object &object : objectList) {
797         const Fortran::semantics::Symbol *symbol = object.id();
798         mlir::Value symVal = converter.getSymbolAddress(*symbol);
799         if (auto declOp = symVal.getDefiningOp<hlfir::DeclareOp>())
800           symVal = declOp.getBase();
801         auto redType = symVal.getType().cast<fir::ReferenceType>();
802         if (!redType.getEleTy().isIntOrIndexOrFloat())
803           TODO(currentLocation, "User Defined Reduction on non-trivial type");
804         decl = createDeclareReduction(
805             firOpBuilder,
806             getReductionName(getRealName(*reductionIntrinsic).ToString(),
807                              firOpBuilder.getKindMap(), redType, isByRef),
808             redId, redType, currentLocation, isByRef);
809         reductionDeclSymbols.push_back(mlir::SymbolRefAttr::get(
810             firOpBuilder.getContext(), decl.getSymName()));
811       }
812     }
813   }
814 }
815 
816 const Fortran::semantics::SourceName
817 ReductionProcessor::getRealName(const Fortran::semantics::Symbol *symbol) {
818   return symbol->GetUltimate().name();
819 }
820 
821 const Fortran::semantics::SourceName
822 ReductionProcessor::getRealName(const omp::clause::ProcedureDesignator &pd) {
823   return getRealName(pd.v.id());
824 }
825 
826 int ReductionProcessor::getOperationIdentity(ReductionIdentifier redId,
827                                              mlir::Location loc) {
828   switch (redId) {
829   case ReductionIdentifier::ADD:
830   case ReductionIdentifier::OR:
831   case ReductionIdentifier::NEQV:
832     return 0;
833   case ReductionIdentifier::MULTIPLY:
834   case ReductionIdentifier::AND:
835   case ReductionIdentifier::EQV:
836     return 1;
837   default:
838     TODO(loc, "Reduction of some intrinsic operators is not supported");
839   }
840 }
841 
842 } // namespace omp
843 } // namespace lower
844 } // namespace Fortran
845