xref: /llvm-project/flang/lib/Lower/OpenMP/ReductionProcessor.cpp (revision f2027a9388728094d84837fc0fdd2e0325362e51) 
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.sym()).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   semantics::Symbol *sym = pd.v.sym();
76   if (!sym->GetUltimate().attrs().test(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 = mlir::dyn_cast<mlir::FloatType>(type)) {
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 = mlir::dyn_cast<mlir::FloatType>(type)) {
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           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 (mlir::isa<mlir::FloatType>(type))
192       return builder.create<mlir::arith::ConstantOp>(
193           loc, type,
194           builder.getFloatAttr(type, (double)getOperationIdentity(redId, loc)));
195 
196     if (mlir::isa<fir::LogicalType>(type)) {
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 /// Generate a fir::ShapeShift op describing the provided boxed array.
299 static fir::ShapeShiftOp getShapeShift(fir::FirOpBuilder &builder,
300                                        mlir::Location loc, mlir::Value box) {
301   fir::SequenceType sequenceType = mlir::cast<fir::SequenceType>(
302       hlfir::getFortranElementOrSequenceType(box.getType()));
303   const unsigned rank = sequenceType.getDimension();
304   llvm::SmallVector<mlir::Value> lbAndExtents;
305   lbAndExtents.reserve(rank * 2);
306 
307   mlir::Type idxTy = builder.getIndexType();
308   for (unsigned i = 0; i < rank; ++i) {
309     // TODO: ideally we want to hoist box reads out of the critical section.
310     // We could do this by having box dimensions in block arguments like
311     // OpenACC does
312     mlir::Value dim = builder.createIntegerConstant(loc, idxTy, i);
313     auto dimInfo =
314         builder.create<fir::BoxDimsOp>(loc, idxTy, idxTy, idxTy, box, dim);
315     lbAndExtents.push_back(dimInfo.getLowerBound());
316     lbAndExtents.push_back(dimInfo.getExtent());
317   }
318 
319   auto shapeShiftTy = fir::ShapeShiftType::get(builder.getContext(), rank);
320   auto shapeShift =
321       builder.create<fir::ShapeShiftOp>(loc, shapeShiftTy, lbAndExtents);
322   return shapeShift;
323 }
324 
325 /// Create reduction combiner region for reduction variables which are boxed
326 /// arrays
327 static void genBoxCombiner(fir::FirOpBuilder &builder, mlir::Location loc,
328                            ReductionProcessor::ReductionIdentifier redId,
329                            fir::BaseBoxType boxTy, mlir::Value lhs,
330                            mlir::Value rhs) {
331   fir::SequenceType seqTy = mlir::dyn_cast_or_null<fir::SequenceType>(
332       fir::unwrapRefType(boxTy.getEleTy()));
333   fir::HeapType heapTy =
334       mlir::dyn_cast_or_null<fir::HeapType>(boxTy.getEleTy());
335   fir::PointerType ptrTy =
336       mlir::dyn_cast_or_null<fir::PointerType>(boxTy.getEleTy());
337   if ((!seqTy || seqTy.hasUnknownShape()) && !heapTy && !ptrTy)
338     TODO(loc, "Unsupported boxed type in OpenMP reduction");
339 
340   // load fir.ref<fir.box<...>>
341   mlir::Value lhsAddr = lhs;
342   lhs = builder.create<fir::LoadOp>(loc, lhs);
343   rhs = builder.create<fir::LoadOp>(loc, rhs);
344 
345   if ((heapTy || ptrTy) && !seqTy) {
346     // get box contents (heap pointers)
347     lhs = builder.create<fir::BoxAddrOp>(loc, lhs);
348     rhs = builder.create<fir::BoxAddrOp>(loc, rhs);
349     mlir::Value lhsValAddr = lhs;
350 
351     // load heap pointers
352     lhs = builder.create<fir::LoadOp>(loc, lhs);
353     rhs = builder.create<fir::LoadOp>(loc, rhs);
354 
355     mlir::Type eleTy = heapTy ? heapTy.getEleTy() : ptrTy.getEleTy();
356 
357     mlir::Value result = ReductionProcessor::createScalarCombiner(
358         builder, loc, redId, eleTy, lhs, rhs);
359     builder.create<fir::StoreOp>(loc, result, lhsValAddr);
360     builder.create<mlir::omp::YieldOp>(loc, lhsAddr);
361     return;
362   }
363 
364   fir::ShapeShiftOp shapeShift = getShapeShift(builder, loc, lhs);
365 
366   // Iterate over array elements, applying the equivalent scalar reduction:
367 
368   // F2018 5.4.10.2: Unallocated allocatable variables may not be referenced
369   // and so no null check is needed here before indexing into the (possibly
370   // allocatable) arrays.
371 
372   // A hlfir::elemental here gets inlined with a temporary so create the
373   // loop nest directly.
374   // This function already controls all of the code in this region so we
375   // know this won't miss any opportuinties for clever elemental inlining
376   hlfir::LoopNest nest = hlfir::genLoopNest(
377       loc, builder, shapeShift.getExtents(), /*isUnordered=*/true);
378   builder.setInsertionPointToStart(nest.innerLoop.getBody());
379   mlir::Type refTy = fir::ReferenceType::get(seqTy.getEleTy());
380   auto lhsEleAddr = builder.create<fir::ArrayCoorOp>(
381       loc, refTy, lhs, shapeShift, /*slice=*/mlir::Value{},
382       nest.oneBasedIndices, /*typeparms=*/mlir::ValueRange{});
383   auto rhsEleAddr = builder.create<fir::ArrayCoorOp>(
384       loc, refTy, rhs, shapeShift, /*slice=*/mlir::Value{},
385       nest.oneBasedIndices, /*typeparms=*/mlir::ValueRange{});
386   auto lhsEle = builder.create<fir::LoadOp>(loc, lhsEleAddr);
387   auto rhsEle = builder.create<fir::LoadOp>(loc, rhsEleAddr);
388   mlir::Value scalarReduction = ReductionProcessor::createScalarCombiner(
389       builder, loc, redId, refTy, lhsEle, rhsEle);
390   builder.create<fir::StoreOp>(loc, scalarReduction, lhsEleAddr);
391 
392   builder.setInsertionPointAfter(nest.outerLoop);
393   builder.create<mlir::omp::YieldOp>(loc, lhsAddr);
394 }
395 
396 // generate combiner region for reduction operations
397 static void genCombiner(fir::FirOpBuilder &builder, mlir::Location loc,
398                         ReductionProcessor::ReductionIdentifier redId,
399                         mlir::Type ty, mlir::Value lhs, mlir::Value rhs,
400                         bool isByRef) {
401   ty = fir::unwrapRefType(ty);
402 
403   if (fir::isa_trivial(ty)) {
404     mlir::Value lhsLoaded = builder.loadIfRef(loc, lhs);
405     mlir::Value rhsLoaded = builder.loadIfRef(loc, rhs);
406 
407     mlir::Value result = ReductionProcessor::createScalarCombiner(
408         builder, loc, redId, ty, lhsLoaded, rhsLoaded);
409     if (isByRef) {
410       builder.create<fir::StoreOp>(loc, result, lhs);
411       builder.create<mlir::omp::YieldOp>(loc, lhs);
412     } else {
413       builder.create<mlir::omp::YieldOp>(loc, result);
414     }
415     return;
416   }
417   // all arrays should have been boxed
418   if (auto boxTy = mlir::dyn_cast<fir::BaseBoxType>(ty)) {
419     genBoxCombiner(builder, loc, redId, boxTy, lhs, rhs);
420     return;
421   }
422 
423   TODO(loc, "OpenMP genCombiner for unsupported reduction variable type");
424 }
425 
426 static void
427 createReductionCleanupRegion(fir::FirOpBuilder &builder, mlir::Location loc,
428                              mlir::omp::DeclareReductionOp &reductionDecl) {
429   mlir::Type redTy = reductionDecl.getType();
430 
431   mlir::Region &cleanupRegion = reductionDecl.getCleanupRegion();
432   assert(cleanupRegion.empty());
433   mlir::Block *block =
434       builder.createBlock(&cleanupRegion, cleanupRegion.end(), {redTy}, {loc});
435   builder.setInsertionPointToEnd(block);
436 
437   auto typeError = [loc]() {
438     fir::emitFatalError(loc,
439                         "Attempt to create an omp reduction cleanup region "
440                         "for a type that wasn't allocated",
441                         /*genCrashDiag=*/true);
442   };
443 
444   mlir::Type valTy = fir::unwrapRefType(redTy);
445   if (auto boxTy = mlir::dyn_cast_or_null<fir::BaseBoxType>(valTy)) {
446     if (!mlir::isa<fir::HeapType, fir::PointerType>(boxTy.getEleTy())) {
447       mlir::Type innerTy = fir::extractSequenceType(boxTy);
448       if (!mlir::isa<fir::SequenceType>(innerTy))
449         typeError();
450     }
451 
452     mlir::Value arg = block->getArgument(0);
453     arg = builder.loadIfRef(loc, arg);
454     assert(mlir::isa<fir::BaseBoxType>(arg.getType()));
455 
456     // Deallocate box
457     // The FIR type system doesn't nesecarrily know that this is a mutable box
458     // if we allocated the thread local array on the heap to avoid looped stack
459     // allocations.
460     mlir::Value addr =
461         hlfir::genVariableRawAddress(loc, builder, hlfir::Entity{arg});
462     mlir::Value isAllocated = builder.genIsNotNullAddr(loc, addr);
463     fir::IfOp ifOp =
464         builder.create<fir::IfOp>(loc, isAllocated, /*withElseRegion=*/false);
465     builder.setInsertionPointToStart(&ifOp.getThenRegion().front());
466 
467     mlir::Value cast = builder.createConvert(
468         loc, fir::HeapType::get(fir::dyn_cast_ptrEleTy(addr.getType())), addr);
469     builder.create<fir::FreeMemOp>(loc, cast);
470 
471     builder.setInsertionPointAfter(ifOp);
472     builder.create<mlir::omp::YieldOp>(loc);
473     return;
474   }
475 
476   typeError();
477 }
478 
479 // like fir::unwrapSeqOrBoxedSeqType except it also works for non-sequence boxes
480 static mlir::Type unwrapSeqOrBoxedType(mlir::Type ty) {
481   if (auto seqTy = mlir::dyn_cast<fir::SequenceType>(ty))
482     return seqTy.getEleTy();
483   if (auto boxTy = mlir::dyn_cast<fir::BaseBoxType>(ty)) {
484     auto eleTy = fir::unwrapRefType(boxTy.getEleTy());
485     if (auto seqTy = mlir::dyn_cast<fir::SequenceType>(eleTy))
486       return seqTy.getEleTy();
487     return eleTy;
488   }
489   return ty;
490 }
491 
492 static void createReductionAllocAndInitRegions(
493     fir::FirOpBuilder &builder, mlir::Location loc,
494     mlir::omp::DeclareReductionOp &reductionDecl,
495     const ReductionProcessor::ReductionIdentifier redId, mlir::Type type,
496     bool isByRef) {
497   auto yield = [&](mlir::Value ret) {
498     builder.create<mlir::omp::YieldOp>(loc, ret);
499   };
500 
501   mlir::Block *allocBlock = nullptr;
502   mlir::Block *initBlock = nullptr;
503   if (isByRef) {
504     allocBlock =
505         builder.createBlock(&reductionDecl.getAllocRegion(),
506                             reductionDecl.getAllocRegion().end(), {}, {});
507     initBlock = builder.createBlock(&reductionDecl.getInitializerRegion(),
508                                     reductionDecl.getInitializerRegion().end(),
509                                     {type, type}, {loc, loc});
510   } else {
511     initBlock = builder.createBlock(&reductionDecl.getInitializerRegion(),
512                                     reductionDecl.getInitializerRegion().end(),
513                                     {type}, {loc});
514   }
515 
516   mlir::Type ty = fir::unwrapRefType(type);
517   builder.setInsertionPointToEnd(initBlock);
518   mlir::Value initValue = ReductionProcessor::getReductionInitValue(
519       loc, unwrapSeqOrBoxedType(ty), redId, builder);
520 
521   if (fir::isa_trivial(ty)) {
522     if (isByRef) {
523       // alloc region
524       {
525         builder.setInsertionPointToEnd(allocBlock);
526         mlir::Value alloca = builder.create<fir::AllocaOp>(loc, ty);
527         yield(alloca);
528       }
529 
530       // init region
531       {
532         builder.setInsertionPointToEnd(initBlock);
533         // block arg is mapped to the alloca yielded from the alloc region
534         mlir::Value alloc = reductionDecl.getInitializerAllocArg();
535         builder.createStoreWithConvert(loc, initValue, alloc);
536         yield(alloc);
537       }
538       return;
539     }
540     // by val
541     yield(initValue);
542     return;
543   }
544 
545   // check if an allocatable box is unallocated. If so, initialize the boxAlloca
546   // to be unallocated e.g.
547   // %box_alloca = fir.alloca !fir.box<!fir.heap<...>>
548   // %addr = fir.box_addr %box
549   // if (%addr == 0) {
550   //   %nullbox = fir.embox %addr
551   //   fir.store %nullbox to %box_alloca
552   // } else {
553   //   // ...
554   //   fir.store %something to %box_alloca
555   // }
556   // omp.yield %box_alloca
557   mlir::Value moldArg =
558       builder.loadIfRef(loc, reductionDecl.getInitializerMoldArg());
559   auto handleNullAllocatable = [&](mlir::Value boxAlloca) -> fir::IfOp {
560     mlir::Value addr = builder.create<fir::BoxAddrOp>(loc, moldArg);
561     mlir::Value isNotAllocated = builder.genIsNullAddr(loc, addr);
562     fir::IfOp ifOp = builder.create<fir::IfOp>(loc, isNotAllocated,
563                                                /*withElseRegion=*/true);
564     builder.setInsertionPointToStart(&ifOp.getThenRegion().front());
565     // just embox the null address and return
566     mlir::Value nullBox = builder.create<fir::EmboxOp>(loc, ty, addr);
567     builder.create<fir::StoreOp>(loc, nullBox, boxAlloca);
568     return ifOp;
569   };
570 
571   // all arrays are boxed
572   if (auto boxTy = mlir::dyn_cast_or_null<fir::BaseBoxType>(ty)) {
573     assert(isByRef && "passing boxes by value is unsupported");
574     bool isAllocatableOrPointer =
575         mlir::isa<fir::HeapType, fir::PointerType>(boxTy.getEleTy());
576 
577     // alloc region
578     {
579       builder.setInsertionPointToEnd(allocBlock);
580       mlir::Value boxAlloca = builder.create<fir::AllocaOp>(loc, ty);
581       yield(boxAlloca);
582     }
583 
584     // init region
585     builder.setInsertionPointToEnd(initBlock);
586     mlir::Value boxAlloca = reductionDecl.getInitializerAllocArg();
587     mlir::Type innerTy = fir::unwrapRefType(boxTy.getEleTy());
588     if (fir::isa_trivial(innerTy)) {
589       // boxed non-sequence value e.g. !fir.box<!fir.heap<i32>>
590       if (!isAllocatableOrPointer)
591         TODO(loc, "Reduction of non-allocatable trivial typed box");
592 
593       fir::IfOp ifUnallocated = handleNullAllocatable(boxAlloca);
594 
595       builder.setInsertionPointToStart(&ifUnallocated.getElseRegion().front());
596       mlir::Value valAlloc = builder.create<fir::AllocMemOp>(loc, innerTy);
597       builder.createStoreWithConvert(loc, initValue, valAlloc);
598       mlir::Value box = builder.create<fir::EmboxOp>(loc, ty, valAlloc);
599       builder.create<fir::StoreOp>(loc, box, boxAlloca);
600 
601       auto insPt = builder.saveInsertionPoint();
602       createReductionCleanupRegion(builder, loc, reductionDecl);
603       builder.restoreInsertionPoint(insPt);
604       builder.setInsertionPointAfter(ifUnallocated);
605       yield(boxAlloca);
606       return;
607     }
608     innerTy = fir::extractSequenceType(boxTy);
609     if (!mlir::isa<fir::SequenceType>(innerTy))
610       TODO(loc, "Unsupported boxed type for reduction");
611 
612     fir::IfOp ifUnallocated{nullptr};
613     if (isAllocatableOrPointer) {
614       ifUnallocated = handleNullAllocatable(boxAlloca);
615       builder.setInsertionPointToStart(&ifUnallocated.getElseRegion().front());
616     }
617 
618     // Create the private copy from the initial fir.box:
619     mlir::Value loadedBox = builder.loadIfRef(loc, moldArg);
620     hlfir::Entity source = hlfir::Entity{loadedBox};
621 
622     // Allocating on the heap in case the whole reduction is nested inside of a
623     // loop
624     // TODO: compare performance here to using allocas - this could be made to
625     // work by inserting stacksave/stackrestore around the reduction in
626     // openmpirbuilder
627     auto [temp, needsDealloc] = createTempFromMold(loc, builder, source);
628     // if needsDealloc isn't statically false, add cleanup region. Always
629     // do this for allocatable boxes because they might have been re-allocated
630     // in the body of the loop/parallel region
631 
632     std::optional<int64_t> cstNeedsDealloc =
633         fir::getIntIfConstant(needsDealloc);
634     assert(cstNeedsDealloc.has_value() &&
635            "createTempFromMold decides this statically");
636     if (cstNeedsDealloc.has_value() && *cstNeedsDealloc != false) {
637       mlir::OpBuilder::InsertionGuard guard(builder);
638       createReductionCleanupRegion(builder, loc, reductionDecl);
639     } else {
640       assert(!isAllocatableOrPointer &&
641              "Pointer-like arrays must be heap allocated");
642     }
643 
644     // Put the temporary inside of a box:
645     // hlfir::genVariableBox doesn't handle non-default lower bounds
646     mlir::Value box;
647     fir::ShapeShiftOp shapeShift = getShapeShift(builder, loc, loadedBox);
648     mlir::Type boxType = loadedBox.getType();
649     if (mlir::isa<fir::BaseBoxType>(temp.getType()))
650       // the box created by the declare form createTempFromMold is missing lower
651       // bounds info
652       box = builder.create<fir::ReboxOp>(loc, boxType, temp, shapeShift,
653                                          /*shift=*/mlir::Value{});
654     else
655       box = builder.create<fir::EmboxOp>(
656           loc, boxType, temp, shapeShift,
657           /*slice=*/mlir::Value{},
658           /*typeParams=*/llvm::ArrayRef<mlir::Value>{});
659 
660     builder.create<hlfir::AssignOp>(loc, initValue, box);
661     builder.create<fir::StoreOp>(loc, box, boxAlloca);
662     if (ifUnallocated)
663       builder.setInsertionPointAfter(ifUnallocated);
664     yield(boxAlloca);
665     return;
666   }
667 
668   TODO(loc, "createReductionInitRegion for unsupported type");
669 }
670 
671 mlir::omp::DeclareReductionOp ReductionProcessor::createDeclareReduction(
672     fir::FirOpBuilder &builder, llvm::StringRef reductionOpName,
673     const ReductionIdentifier redId, mlir::Type type, mlir::Location loc,
674     bool isByRef) {
675   mlir::OpBuilder::InsertionGuard guard(builder);
676   mlir::ModuleOp module = builder.getModule();
677 
678   assert(!reductionOpName.empty());
679 
680   auto decl =
681       module.lookupSymbol<mlir::omp::DeclareReductionOp>(reductionOpName);
682   if (decl)
683     return decl;
684 
685   mlir::OpBuilder modBuilder(module.getBodyRegion());
686   mlir::Type valTy = fir::unwrapRefType(type);
687   if (!isByRef)
688     type = valTy;
689 
690   decl = modBuilder.create<mlir::omp::DeclareReductionOp>(loc, reductionOpName,
691                                                           type);
692   createReductionAllocAndInitRegions(builder, loc, decl, redId, type, isByRef);
693 
694   builder.createBlock(&decl.getReductionRegion(),
695                       decl.getReductionRegion().end(), {type, type},
696                       {loc, loc});
697 
698   builder.setInsertionPointToEnd(&decl.getReductionRegion().back());
699   mlir::Value op1 = decl.getReductionRegion().front().getArgument(0);
700   mlir::Value op2 = decl.getReductionRegion().front().getArgument(1);
701   genCombiner(builder, loc, redId, type, op1, op2, isByRef);
702 
703   return decl;
704 }
705 
706 static bool doReductionByRef(mlir::Value reductionVar) {
707   if (forceByrefReduction)
708     return true;
709 
710   if (auto declare =
711           mlir::dyn_cast<hlfir::DeclareOp>(reductionVar.getDefiningOp()))
712     reductionVar = declare.getMemref();
713 
714   if (!fir::isa_trivial(fir::unwrapRefType(reductionVar.getType())))
715     return true;
716 
717   return false;
718 }
719 
720 void ReductionProcessor::addDeclareReduction(
721     mlir::Location currentLocation, lower::AbstractConverter &converter,
722     const omp::clause::Reduction &reduction,
723     llvm::SmallVectorImpl<mlir::Value> &reductionVars,
724     llvm::SmallVectorImpl<bool> &reduceVarByRef,
725     llvm::SmallVectorImpl<mlir::Attribute> &reductionDeclSymbols,
726     llvm::SmallVectorImpl<const semantics::Symbol *> *reductionSymbols) {
727   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
728 
729   if (std::get<std::optional<omp::clause::Reduction::ReductionModifier>>(
730           reduction.t))
731     TODO(currentLocation, "Reduction modifiers are not supported");
732 
733   mlir::omp::DeclareReductionOp decl;
734   const auto &redOperatorList{
735       std::get<omp::clause::Reduction::ReductionIdentifiers>(reduction.t)};
736   assert(redOperatorList.size() == 1 && "Expecting single operator");
737   const auto &redOperator = redOperatorList.front();
738   const auto &objectList{std::get<omp::ObjectList>(reduction.t)};
739 
740   if (!std::holds_alternative<omp::clause::DefinedOperator>(redOperator.u)) {
741     if (const auto *reductionIntrinsic =
742             std::get_if<omp::clause::ProcedureDesignator>(&redOperator.u)) {
743       if (!ReductionProcessor::supportedIntrinsicProcReduction(
744               *reductionIntrinsic)) {
745         return;
746       }
747     } else {
748       return;
749     }
750   }
751 
752   // Reduction variable processing common to both intrinsic operators and
753   // procedure designators
754   fir::FirOpBuilder &builder = converter.getFirOpBuilder();
755   for (const Object &object : objectList) {
756     const semantics::Symbol *symbol = object.sym();
757     if (reductionSymbols)
758       reductionSymbols->push_back(symbol);
759     mlir::Value symVal = converter.getSymbolAddress(*symbol);
760     mlir::Type eleType;
761     auto refType = mlir::dyn_cast_or_null<fir::ReferenceType>(symVal.getType());
762     if (refType)
763       eleType = refType.getEleTy();
764     else
765       eleType = symVal.getType();
766 
767     // all arrays must be boxed so that we have convenient access to all the
768     // information needed to iterate over the array
769     if (mlir::isa<fir::SequenceType>(eleType)) {
770       // For Host associated symbols, use `SymbolBox` instead
771       lower::SymbolBox symBox = converter.lookupOneLevelUpSymbol(*symbol);
772       hlfir::Entity entity{symBox.getAddr()};
773       entity = genVariableBox(currentLocation, builder, entity);
774       mlir::Value box = entity.getBase();
775 
776       // Always pass the box by reference so that the OpenMP dialect
777       // verifiers don't need to know anything about fir.box
778       auto alloca =
779           builder.create<fir::AllocaOp>(currentLocation, box.getType());
780       builder.create<fir::StoreOp>(currentLocation, box, alloca);
781 
782       symVal = alloca;
783     } else if (mlir::isa<fir::BaseBoxType>(symVal.getType())) {
784       // boxed arrays are passed as values not by reference. Unfortunately,
785       // we can't pass a box by value to omp.redution_declare, so turn it
786       // into a reference
787 
788       auto alloca =
789           builder.create<fir::AllocaOp>(currentLocation, symVal.getType());
790       builder.create<fir::StoreOp>(currentLocation, symVal, alloca);
791       symVal = alloca;
792     } else if (auto declOp = symVal.getDefiningOp<hlfir::DeclareOp>()) {
793       symVal = declOp.getBase();
794     }
795 
796     // this isn't the same as the by-val and by-ref passing later in the
797     // pipeline. Both styles assume that the variable is a reference at
798     // this point
799     assert(mlir::isa<fir::ReferenceType>(symVal.getType()) &&
800            "reduction input var is a reference");
801 
802     reductionVars.push_back(symVal);
803     reduceVarByRef.push_back(doReductionByRef(symVal));
804   }
805 
806   for (auto [symVal, isByRef] : llvm::zip(reductionVars, reduceVarByRef)) {
807     auto redType = mlir::cast<fir::ReferenceType>(symVal.getType());
808     const auto &kindMap = firOpBuilder.getKindMap();
809     std::string reductionName;
810     ReductionIdentifier redId;
811     mlir::Type redNameTy = redType;
812     if (mlir::isa<fir::LogicalType>(redType.getEleTy()))
813       redNameTy = builder.getI1Type();
814 
815     if (const auto &redDefinedOp =
816             std::get_if<omp::clause::DefinedOperator>(&redOperator.u)) {
817       const auto &intrinsicOp{
818           std::get<omp::clause::DefinedOperator::IntrinsicOperator>(
819               redDefinedOp->u)};
820       redId = getReductionType(intrinsicOp);
821       switch (redId) {
822       case ReductionIdentifier::ADD:
823       case ReductionIdentifier::MULTIPLY:
824       case ReductionIdentifier::AND:
825       case ReductionIdentifier::EQV:
826       case ReductionIdentifier::OR:
827       case ReductionIdentifier::NEQV:
828         break;
829       default:
830         TODO(currentLocation,
831              "Reduction of some intrinsic operators is not supported");
832         break;
833       }
834 
835       reductionName =
836           getReductionName(intrinsicOp, kindMap, redNameTy, isByRef);
837     } else if (const auto *reductionIntrinsic =
838                    std::get_if<omp::clause::ProcedureDesignator>(
839                        &redOperator.u)) {
840       if (!ReductionProcessor::supportedIntrinsicProcReduction(
841               *reductionIntrinsic)) {
842         TODO(currentLocation, "Unsupported intrinsic proc reduction");
843       }
844       redId = getReductionType(*reductionIntrinsic);
845       reductionName =
846           getReductionName(getRealName(*reductionIntrinsic).ToString(), kindMap,
847                            redNameTy, isByRef);
848     } else {
849       TODO(currentLocation, "Unexpected reduction type");
850     }
851 
852     decl = createDeclareReduction(firOpBuilder, reductionName, redId, redType,
853                                   currentLocation, isByRef);
854     reductionDeclSymbols.push_back(
855         mlir::SymbolRefAttr::get(firOpBuilder.getContext(), decl.getSymName()));
856   }
857 }
858 
859 const semantics::SourceName
860 ReductionProcessor::getRealName(const semantics::Symbol *symbol) {
861   return symbol->GetUltimate().name();
862 }
863 
864 const semantics::SourceName
865 ReductionProcessor::getRealName(const omp::clause::ProcedureDesignator &pd) {
866   return getRealName(pd.v.sym());
867 }
868 
869 int ReductionProcessor::getOperationIdentity(ReductionIdentifier redId,
870                                              mlir::Location loc) {
871   switch (redId) {
872   case ReductionIdentifier::ADD:
873   case ReductionIdentifier::OR:
874   case ReductionIdentifier::NEQV:
875     return 0;
876   case ReductionIdentifier::MULTIPLY:
877   case ReductionIdentifier::AND:
878   case ReductionIdentifier::EQV:
879     return 1;
880   default:
881     TODO(loc, "Reduction of some intrinsic operators is not supported");
882   }
883 }
884 
885 } // namespace omp
886 } // namespace lower
887 } // namespace Fortran
888