xref: /llvm-project/flang/lib/Lower/OpenMP/ReductionProcessor.cpp (revision fac349a169976f822fb27f03e623fa0d28aec1f3) 
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 = 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           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 (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   if ((!seqTy || seqTy.hasUnknownShape()) && !heapTy)
336     TODO(loc, "Unsupported boxed type in OpenMP reduction");
337 
338   // load fir.ref<fir.box<...>>
339   mlir::Value lhsAddr = lhs;
340   lhs = builder.create<fir::LoadOp>(loc, lhs);
341   rhs = builder.create<fir::LoadOp>(loc, rhs);
342 
343   if (heapTy && !seqTy) {
344     // get box contents (heap pointers)
345     lhs = builder.create<fir::BoxAddrOp>(loc, lhs);
346     rhs = builder.create<fir::BoxAddrOp>(loc, rhs);
347     mlir::Value lhsValAddr = lhs;
348 
349     // load heap pointers
350     lhs = builder.create<fir::LoadOp>(loc, lhs);
351     rhs = builder.create<fir::LoadOp>(loc, rhs);
352 
353     mlir::Value result = ReductionProcessor::createScalarCombiner(
354         builder, loc, redId, heapTy.getEleTy(), lhs, rhs);
355     builder.create<fir::StoreOp>(loc, result, lhsValAddr);
356     builder.create<mlir::omp::YieldOp>(loc, lhsAddr);
357     return;
358   }
359 
360   fir::ShapeShiftOp shapeShift = getShapeShift(builder, loc, lhs);
361 
362   // Iterate over array elements, applying the equivalent scalar reduction:
363 
364   // F2018 5.4.10.2: Unallocated allocatable variables may not be referenced
365   // and so no null check is needed here before indexing into the (possibly
366   // allocatable) arrays.
367 
368   // A hlfir::elemental here gets inlined with a temporary so create the
369   // loop nest directly.
370   // This function already controls all of the code in this region so we
371   // know this won't miss any opportuinties for clever elemental inlining
372   hlfir::LoopNest nest = hlfir::genLoopNest(
373       loc, builder, shapeShift.getExtents(), /*isUnordered=*/true);
374   builder.setInsertionPointToStart(nest.innerLoop.getBody());
375   mlir::Type refTy = fir::ReferenceType::get(seqTy.getEleTy());
376   auto lhsEleAddr = builder.create<fir::ArrayCoorOp>(
377       loc, refTy, lhs, shapeShift, /*slice=*/mlir::Value{},
378       nest.oneBasedIndices, /*typeparms=*/mlir::ValueRange{});
379   auto rhsEleAddr = builder.create<fir::ArrayCoorOp>(
380       loc, refTy, rhs, shapeShift, /*slice=*/mlir::Value{},
381       nest.oneBasedIndices, /*typeparms=*/mlir::ValueRange{});
382   auto lhsEle = builder.create<fir::LoadOp>(loc, lhsEleAddr);
383   auto rhsEle = builder.create<fir::LoadOp>(loc, rhsEleAddr);
384   mlir::Value scalarReduction = ReductionProcessor::createScalarCombiner(
385       builder, loc, redId, refTy, lhsEle, rhsEle);
386   builder.create<fir::StoreOp>(loc, scalarReduction, lhsEleAddr);
387 
388   builder.setInsertionPointAfter(nest.outerLoop);
389   builder.create<mlir::omp::YieldOp>(loc, lhsAddr);
390 }
391 
392 // generate combiner region for reduction operations
393 static void genCombiner(fir::FirOpBuilder &builder, mlir::Location loc,
394                         ReductionProcessor::ReductionIdentifier redId,
395                         mlir::Type ty, mlir::Value lhs, mlir::Value rhs,
396                         bool isByRef) {
397   ty = fir::unwrapRefType(ty);
398 
399   if (fir::isa_trivial(ty)) {
400     mlir::Value lhsLoaded = builder.loadIfRef(loc, lhs);
401     mlir::Value rhsLoaded = builder.loadIfRef(loc, rhs);
402 
403     mlir::Value result = ReductionProcessor::createScalarCombiner(
404         builder, loc, redId, ty, lhsLoaded, rhsLoaded);
405     if (isByRef) {
406       builder.create<fir::StoreOp>(loc, result, lhs);
407       builder.create<mlir::omp::YieldOp>(loc, lhs);
408     } else {
409       builder.create<mlir::omp::YieldOp>(loc, result);
410     }
411     return;
412   }
413   // all arrays should have been boxed
414   if (auto boxTy = mlir::dyn_cast<fir::BaseBoxType>(ty)) {
415     genBoxCombiner(builder, loc, redId, boxTy, lhs, rhs);
416     return;
417   }
418 
419   TODO(loc, "OpenMP genCombiner for unsupported reduction variable type");
420 }
421 
422 static void
423 createReductionCleanupRegion(fir::FirOpBuilder &builder, mlir::Location loc,
424                              mlir::omp::DeclareReductionOp &reductionDecl) {
425   mlir::Type redTy = reductionDecl.getType();
426 
427   mlir::Region &cleanupRegion = reductionDecl.getCleanupRegion();
428   assert(cleanupRegion.empty());
429   mlir::Block *block =
430       builder.createBlock(&cleanupRegion, cleanupRegion.end(), {redTy}, {loc});
431   builder.setInsertionPointToEnd(block);
432 
433   auto typeError = [loc]() {
434     fir::emitFatalError(loc,
435                         "Attempt to create an omp reduction cleanup region "
436                         "for a type that wasn't allocated",
437                         /*genCrashDiag=*/true);
438   };
439 
440   mlir::Type valTy = fir::unwrapRefType(redTy);
441   if (auto boxTy = mlir::dyn_cast_or_null<fir::BaseBoxType>(valTy)) {
442     if (!mlir::isa<fir::HeapType>(boxTy.getEleTy())) {
443       mlir::Type innerTy = fir::extractSequenceType(boxTy);
444       if (!mlir::isa<fir::SequenceType>(innerTy))
445         typeError();
446     }
447 
448     mlir::Value arg = block->getArgument(0);
449     arg = builder.loadIfRef(loc, arg);
450     assert(mlir::isa<fir::BaseBoxType>(arg.getType()));
451 
452     // Deallocate box
453     // The FIR type system doesn't nesecarrily know that this is a mutable box
454     // if we allocated the thread local array on the heap to avoid looped stack
455     // allocations.
456     mlir::Value addr =
457         hlfir::genVariableRawAddress(loc, builder, hlfir::Entity{arg});
458     mlir::Value isAllocated = builder.genIsNotNullAddr(loc, addr);
459     fir::IfOp ifOp =
460         builder.create<fir::IfOp>(loc, isAllocated, /*withElseRegion=*/false);
461     builder.setInsertionPointToStart(&ifOp.getThenRegion().front());
462 
463     mlir::Value cast = builder.createConvert(
464         loc, fir::HeapType::get(fir::dyn_cast_ptrEleTy(addr.getType())), addr);
465     builder.create<fir::FreeMemOp>(loc, cast);
466 
467     builder.setInsertionPointAfter(ifOp);
468     builder.create<mlir::omp::YieldOp>(loc);
469     return;
470   }
471 
472   typeError();
473 }
474 
475 // like fir::unwrapSeqOrBoxedSeqType except it also works for non-sequence boxes
476 static mlir::Type unwrapSeqOrBoxedType(mlir::Type ty) {
477   if (auto seqTy = mlir::dyn_cast<fir::SequenceType>(ty))
478     return seqTy.getEleTy();
479   if (auto boxTy = mlir::dyn_cast<fir::BaseBoxType>(ty)) {
480     auto eleTy = fir::unwrapRefType(boxTy.getEleTy());
481     if (auto seqTy = mlir::dyn_cast<fir::SequenceType>(eleTy))
482       return seqTy.getEleTy();
483     return eleTy;
484   }
485   return ty;
486 }
487 
488 static mlir::Value
489 createReductionInitRegion(fir::FirOpBuilder &builder, mlir::Location loc,
490                           mlir::omp::DeclareReductionOp &reductionDecl,
491                           const ReductionProcessor::ReductionIdentifier redId,
492                           mlir::Type type, bool isByRef) {
493   mlir::Type ty = fir::unwrapRefType(type);
494   mlir::Value initValue = ReductionProcessor::getReductionInitValue(
495       loc, unwrapSeqOrBoxedType(ty), redId, builder);
496 
497   if (fir::isa_trivial(ty)) {
498     if (isByRef) {
499       mlir::Value alloca = builder.create<fir::AllocaOp>(loc, ty);
500       builder.createStoreWithConvert(loc, initValue, alloca);
501       return alloca;
502     }
503     // by val
504     return initValue;
505   }
506 
507   // check if an allocatable box is unallocated. If so, initialize the boxAlloca
508   // to be unallocated e.g.
509   // %box_alloca = fir.alloca !fir.box<!fir.heap<...>>
510   // %addr = fir.box_addr %box
511   // if (%addr == 0) {
512   //   %nullbox = fir.embox %addr
513   //   fir.store %nullbox to %box_alloca
514   // } else {
515   //   // ...
516   //   fir.store %something to %box_alloca
517   // }
518   // omp.yield %box_alloca
519   mlir::Value blockArg =
520       builder.loadIfRef(loc, builder.getBlock()->getArgument(0));
521   auto handleNullAllocatable = [&](mlir::Value boxAlloca) -> fir::IfOp {
522     mlir::Value addr = builder.create<fir::BoxAddrOp>(loc, blockArg);
523     mlir::Value isNotAllocated = builder.genIsNullAddr(loc, addr);
524     fir::IfOp ifOp = builder.create<fir::IfOp>(loc, isNotAllocated,
525                                                /*withElseRegion=*/true);
526     builder.setInsertionPointToStart(&ifOp.getThenRegion().front());
527     // just embox the null address and return
528     mlir::Value nullBox = builder.create<fir::EmboxOp>(loc, ty, addr);
529     builder.create<fir::StoreOp>(loc, nullBox, boxAlloca);
530     return ifOp;
531   };
532 
533   // all arrays are boxed
534   if (auto boxTy = mlir::dyn_cast_or_null<fir::BaseBoxType>(ty)) {
535     assert(isByRef && "passing boxes by value is unsupported");
536     bool isAllocatable = mlir::isa<fir::HeapType>(boxTy.getEleTy());
537     mlir::Value boxAlloca = builder.create<fir::AllocaOp>(loc, ty);
538     mlir::Type innerTy = fir::unwrapRefType(boxTy.getEleTy());
539     if (fir::isa_trivial(innerTy)) {
540       // boxed non-sequence value e.g. !fir.box<!fir.heap<i32>>
541       if (!isAllocatable)
542         TODO(loc, "Reduction of non-allocatable trivial typed box");
543 
544       fir::IfOp ifUnallocated = handleNullAllocatable(boxAlloca);
545 
546       builder.setInsertionPointToStart(&ifUnallocated.getElseRegion().front());
547       mlir::Value valAlloc = builder.create<fir::AllocMemOp>(loc, innerTy);
548       builder.createStoreWithConvert(loc, initValue, valAlloc);
549       mlir::Value box = builder.create<fir::EmboxOp>(loc, ty, valAlloc);
550       builder.create<fir::StoreOp>(loc, box, boxAlloca);
551 
552       auto insPt = builder.saveInsertionPoint();
553       createReductionCleanupRegion(builder, loc, reductionDecl);
554       builder.restoreInsertionPoint(insPt);
555       builder.setInsertionPointAfter(ifUnallocated);
556       return boxAlloca;
557     }
558     innerTy = fir::extractSequenceType(boxTy);
559     if (!mlir::isa<fir::SequenceType>(innerTy))
560       TODO(loc, "Unsupported boxed type for reduction");
561 
562     fir::IfOp ifUnallocated{nullptr};
563     if (isAllocatable) {
564       ifUnallocated = handleNullAllocatable(boxAlloca);
565       builder.setInsertionPointToStart(&ifUnallocated.getElseRegion().front());
566     }
567 
568     // Create the private copy from the initial fir.box:
569     mlir::Value loadedBox = builder.loadIfRef(loc, blockArg);
570     hlfir::Entity source = hlfir::Entity{loadedBox};
571 
572     // Allocating on the heap in case the whole reduction is nested inside of a
573     // loop
574     // TODO: compare performance here to using allocas - this could be made to
575     // work by inserting stacksave/stackrestore around the reduction in
576     // openmpirbuilder
577     auto [temp, needsDealloc] = createTempFromMold(loc, builder, source);
578     // if needsDealloc isn't statically false, add cleanup region. Always
579     // do this for allocatable boxes because they might have been re-allocated
580     // in the body of the loop/parallel region
581 
582     std::optional<int64_t> cstNeedsDealloc =
583         fir::getIntIfConstant(needsDealloc);
584     assert(cstNeedsDealloc.has_value() &&
585            "createTempFromMold decides this statically");
586     if (cstNeedsDealloc.has_value() && *cstNeedsDealloc != false) {
587       mlir::OpBuilder::InsertionGuard guard(builder);
588       createReductionCleanupRegion(builder, loc, reductionDecl);
589     } else {
590       assert(!isAllocatable && "Allocatable arrays must be heap allocated");
591     }
592 
593     // Put the temporary inside of a box:
594     // hlfir::genVariableBox doesn't handle non-default lower bounds
595     mlir::Value box;
596     fir::ShapeShiftOp shapeShift = getShapeShift(builder, loc, loadedBox);
597     mlir::Type boxType = loadedBox.getType();
598     if (mlir::isa<fir::BaseBoxType>(temp.getType()))
599       // the box created by the declare form createTempFromMold is missing lower
600       // bounds info
601       box = builder.create<fir::ReboxOp>(loc, boxType, temp, shapeShift,
602                                          /*shift=*/mlir::Value{});
603     else
604       box = builder.create<fir::EmboxOp>(
605           loc, boxType, temp, shapeShift,
606           /*slice=*/mlir::Value{},
607           /*typeParams=*/llvm::ArrayRef<mlir::Value>{});
608 
609     builder.create<hlfir::AssignOp>(loc, initValue, box);
610     builder.create<fir::StoreOp>(loc, box, boxAlloca);
611     if (ifUnallocated)
612       builder.setInsertionPointAfter(ifUnallocated);
613     return boxAlloca;
614   }
615 
616   TODO(loc, "createReductionInitRegion for unsupported type");
617 }
618 
619 mlir::omp::DeclareReductionOp ReductionProcessor::createDeclareReduction(
620     fir::FirOpBuilder &builder, llvm::StringRef reductionOpName,
621     const ReductionIdentifier redId, mlir::Type type, mlir::Location loc,
622     bool isByRef) {
623   mlir::OpBuilder::InsertionGuard guard(builder);
624   mlir::ModuleOp module = builder.getModule();
625 
626   assert(!reductionOpName.empty());
627 
628   auto decl =
629       module.lookupSymbol<mlir::omp::DeclareReductionOp>(reductionOpName);
630   if (decl)
631     return decl;
632 
633   mlir::OpBuilder modBuilder(module.getBodyRegion());
634   mlir::Type valTy = fir::unwrapRefType(type);
635   if (!isByRef)
636     type = valTy;
637 
638   decl = modBuilder.create<mlir::omp::DeclareReductionOp>(loc, reductionOpName,
639                                                           type);
640   builder.createBlock(&decl.getInitializerRegion(),
641                       decl.getInitializerRegion().end(), {type}, {loc});
642   builder.setInsertionPointToEnd(&decl.getInitializerRegion().back());
643 
644   mlir::Value init =
645       createReductionInitRegion(builder, loc, decl, redId, type, isByRef);
646   builder.create<mlir::omp::YieldOp>(loc, init);
647 
648   builder.createBlock(&decl.getReductionRegion(),
649                       decl.getReductionRegion().end(), {type, type},
650                       {loc, loc});
651 
652   builder.setInsertionPointToEnd(&decl.getReductionRegion().back());
653   mlir::Value op1 = decl.getReductionRegion().front().getArgument(0);
654   mlir::Value op2 = decl.getReductionRegion().front().getArgument(1);
655   genCombiner(builder, loc, redId, type, op1, op2, isByRef);
656 
657   return decl;
658 }
659 
660 // TODO: By-ref vs by-val reductions are currently toggled for the whole
661 //       operation (possibly effecting multiple reduction variables).
662 //       This could cause a problem with openmp target reductions because
663 //       by-ref trivial types may not be supported.
664 bool ReductionProcessor::doReductionByRef(
665     const llvm::SmallVectorImpl<mlir::Value> &reductionVars) {
666   if (reductionVars.empty())
667     return false;
668   if (forceByrefReduction)
669     return true;
670 
671   for (mlir::Value reductionVar : reductionVars) {
672     if (auto declare =
673             mlir::dyn_cast<hlfir::DeclareOp>(reductionVar.getDefiningOp()))
674       reductionVar = declare.getMemref();
675 
676     if (!fir::isa_trivial(fir::unwrapRefType(reductionVar.getType())))
677       return true;
678   }
679   return false;
680 }
681 
682 void ReductionProcessor::addDeclareReduction(
683     mlir::Location currentLocation,
684     Fortran::lower::AbstractConverter &converter,
685     const omp::clause::Reduction &reduction,
686     llvm::SmallVectorImpl<mlir::Value> &reductionVars,
687     llvm::SmallVectorImpl<mlir::Attribute> &reductionDeclSymbols,
688     llvm::SmallVectorImpl<const Fortran::semantics::Symbol *>
689         *reductionSymbols) {
690   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
691 
692   if (std::get<std::optional<omp::clause::Reduction::ReductionModifier>>(
693           reduction.t))
694     TODO(currentLocation, "Reduction modifiers are not supported");
695 
696   mlir::omp::DeclareReductionOp decl;
697   const auto &redOperatorList{
698       std::get<omp::clause::Reduction::ReductionIdentifiers>(reduction.t)};
699   assert(redOperatorList.size() == 1 && "Expecting single operator");
700   const auto &redOperator = redOperatorList.front();
701   const auto &objectList{std::get<omp::ObjectList>(reduction.t)};
702 
703   if (!std::holds_alternative<omp::clause::DefinedOperator>(redOperator.u)) {
704     if (const auto *reductionIntrinsic =
705             std::get_if<omp::clause::ProcedureDesignator>(&redOperator.u)) {
706       if (!ReductionProcessor::supportedIntrinsicProcReduction(
707               *reductionIntrinsic)) {
708         return;
709       }
710     } else {
711       return;
712     }
713   }
714 
715   // initial pass to collect all reduction vars so we can figure out if this
716   // should happen byref
717   fir::FirOpBuilder &builder = converter.getFirOpBuilder();
718   for (const Object &object : objectList) {
719     const Fortran::semantics::Symbol *symbol = object.id();
720     if (reductionSymbols)
721       reductionSymbols->push_back(symbol);
722     mlir::Value symVal = converter.getSymbolAddress(*symbol);
723     mlir::Type eleType;
724     auto refType = mlir::dyn_cast_or_null<fir::ReferenceType>(symVal.getType());
725     if (refType)
726       eleType = refType.getEleTy();
727     else
728       eleType = symVal.getType();
729 
730     // all arrays must be boxed so that we have convenient access to all the
731     // information needed to iterate over the array
732     if (mlir::isa<fir::SequenceType>(eleType)) {
733       // For Host associated symbols, use `SymbolBox` instead
734       Fortran::lower::SymbolBox symBox =
735           converter.lookupOneLevelUpSymbol(*symbol);
736       hlfir::Entity entity{symBox.getAddr()};
737       entity = genVariableBox(currentLocation, builder, entity);
738       mlir::Value box = entity.getBase();
739 
740       // Always pass the box by reference so that the OpenMP dialect
741       // verifiers don't need to know anything about fir.box
742       auto alloca =
743           builder.create<fir::AllocaOp>(currentLocation, box.getType());
744       builder.create<fir::StoreOp>(currentLocation, box, alloca);
745 
746       symVal = alloca;
747     } else if (mlir::isa<fir::BaseBoxType>(symVal.getType())) {
748       // boxed arrays are passed as values not by reference. Unfortunately,
749       // we can't pass a box by value to omp.redution_declare, so turn it
750       // into a reference
751 
752       auto alloca =
753           builder.create<fir::AllocaOp>(currentLocation, symVal.getType());
754       builder.create<fir::StoreOp>(currentLocation, symVal, alloca);
755       symVal = alloca;
756     } else if (auto declOp = symVal.getDefiningOp<hlfir::DeclareOp>()) {
757       symVal = declOp.getBase();
758     }
759 
760     // this isn't the same as the by-val and by-ref passing later in the
761     // pipeline. Both styles assume that the variable is a reference at
762     // this point
763     assert(mlir::isa<fir::ReferenceType>(symVal.getType()) &&
764            "reduction input var is a reference");
765 
766     reductionVars.push_back(symVal);
767   }
768   const bool isByRef = doReductionByRef(reductionVars);
769 
770   if (const auto &redDefinedOp =
771           std::get_if<omp::clause::DefinedOperator>(&redOperator.u)) {
772     const auto &intrinsicOp{
773         std::get<omp::clause::DefinedOperator::IntrinsicOperator>(
774             redDefinedOp->u)};
775     ReductionIdentifier redId = getReductionType(intrinsicOp);
776     switch (redId) {
777     case ReductionIdentifier::ADD:
778     case ReductionIdentifier::MULTIPLY:
779     case ReductionIdentifier::AND:
780     case ReductionIdentifier::EQV:
781     case ReductionIdentifier::OR:
782     case ReductionIdentifier::NEQV:
783       break;
784     default:
785       TODO(currentLocation,
786            "Reduction of some intrinsic operators is not supported");
787       break;
788     }
789 
790     for (mlir::Value symVal : reductionVars) {
791       auto redType = mlir::cast<fir::ReferenceType>(symVal.getType());
792       const auto &kindMap = firOpBuilder.getKindMap();
793       if (mlir::isa<fir::LogicalType>(redType.getEleTy()))
794         decl = createDeclareReduction(firOpBuilder,
795                                       getReductionName(intrinsicOp, kindMap,
796                                                        firOpBuilder.getI1Type(),
797                                                        isByRef),
798                                       redId, redType, currentLocation, isByRef);
799       else
800         decl = createDeclareReduction(
801             firOpBuilder,
802             getReductionName(intrinsicOp, kindMap, redType, isByRef), redId,
803             redType, currentLocation, isByRef);
804       reductionDeclSymbols.push_back(mlir::SymbolRefAttr::get(
805           firOpBuilder.getContext(), decl.getSymName()));
806     }
807   } else if (const auto *reductionIntrinsic =
808                  std::get_if<omp::clause::ProcedureDesignator>(
809                      &redOperator.u)) {
810     if (ReductionProcessor::supportedIntrinsicProcReduction(
811             *reductionIntrinsic)) {
812       ReductionProcessor::ReductionIdentifier redId =
813           ReductionProcessor::getReductionType(*reductionIntrinsic);
814       for (const Object &object : objectList) {
815         const Fortran::semantics::Symbol *symbol = object.id();
816         mlir::Value symVal = converter.getSymbolAddress(*symbol);
817         if (auto declOp = symVal.getDefiningOp<hlfir::DeclareOp>())
818           symVal = declOp.getBase();
819         auto redType = mlir::cast<fir::ReferenceType>(symVal.getType());
820         if (!redType.getEleTy().isIntOrIndexOrFloat())
821           TODO(currentLocation, "User Defined Reduction on non-trivial type");
822         decl = createDeclareReduction(
823             firOpBuilder,
824             getReductionName(getRealName(*reductionIntrinsic).ToString(),
825                              firOpBuilder.getKindMap(), redType, isByRef),
826             redId, redType, currentLocation, isByRef);
827         reductionDeclSymbols.push_back(mlir::SymbolRefAttr::get(
828             firOpBuilder.getContext(), decl.getSymName()));
829       }
830     }
831   }
832 }
833 
834 const Fortran::semantics::SourceName
835 ReductionProcessor::getRealName(const Fortran::semantics::Symbol *symbol) {
836   return symbol->GetUltimate().name();
837 }
838 
839 const Fortran::semantics::SourceName
840 ReductionProcessor::getRealName(const omp::clause::ProcedureDesignator &pd) {
841   return getRealName(pd.v.id());
842 }
843 
844 int ReductionProcessor::getOperationIdentity(ReductionIdentifier redId,
845                                              mlir::Location loc) {
846   switch (redId) {
847   case ReductionIdentifier::ADD:
848   case ReductionIdentifier::OR:
849   case ReductionIdentifier::NEQV:
850     return 0;
851   case ReductionIdentifier::MULTIPLY:
852   case ReductionIdentifier::AND:
853   case ReductionIdentifier::EQV:
854     return 1;
855   default:
856     TODO(loc, "Reduction of some intrinsic operators is not supported");
857   }
858 }
859 
860 } // namespace omp
861 } // namespace lower
862 } // namespace Fortran
863