xref: /llvm-project/mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp (revision 039b969b32b64b64123dce30dd28ec4e343d893f)
1 //===- SCFToOpenMP.cpp - Structured Control Flow to OpenMP conversion -----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements a pass to convert scf.parallel operations into OpenMP
10 // parallel loops.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "mlir/Conversion/SCFToOpenMP/SCFToOpenMP.h"
15 #include "../PassDetail.h"
16 #include "mlir/Analysis/SliceAnalysis.h"
17 #include "mlir/Dialect/Affine/Analysis/LoopAnalysis.h"
18 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
19 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
20 #include "mlir/Dialect/MemRef/IR/MemRef.h"
21 #include "mlir/Dialect/OpenMP/OpenMPDialect.h"
22 #include "mlir/Dialect/SCF/IR/SCF.h"
23 #include "mlir/IR/ImplicitLocOpBuilder.h"
24 #include "mlir/IR/SymbolTable.h"
25 #include "mlir/Transforms/DialectConversion.h"
26 
27 using namespace mlir;
28 
29 /// Matches a block containing a "simple" reduction. The expected shape of the
30 /// block is as follows.
31 ///
32 ///   ^bb(%arg0, %arg1):
33 ///     %0 = OpTy(%arg0, %arg1)
34 ///     scf.reduce.return %0
35 template <typename... OpTy>
36 static bool matchSimpleReduction(Block &block) {
37   if (block.empty() || llvm::hasSingleElement(block) ||
38       std::next(block.begin(), 2) != block.end())
39     return false;
40 
41   if (block.getNumArguments() != 2)
42     return false;
43 
44   SmallVector<Operation *, 4> combinerOps;
45   Value reducedVal = matchReduction({block.getArguments()[1]},
46                                     /*redPos=*/0, combinerOps);
47 
48   if (!reducedVal || !reducedVal.isa<BlockArgument>() ||
49       combinerOps.size() != 1)
50     return false;
51 
52   return isa<OpTy...>(combinerOps[0]) &&
53          isa<scf::ReduceReturnOp>(block.back()) &&
54          block.front().getOperands() == block.getArguments();
55 }
56 
57 /// Matches a block containing a select-based min/max reduction. The types of
58 /// select and compare operations are provided as template arguments. The
59 /// comparison predicates suitable for min and max are provided as function
60 /// arguments. If a reduction is matched, `ifMin` will be set if the reduction
61 /// compute the minimum and unset if it computes the maximum, otherwise it
62 /// remains unmodified. The expected shape of the block is as follows.
63 ///
64 ///   ^bb(%arg0, %arg1):
65 ///     %0 = CompareOpTy(<one-of-predicates>, %arg0, %arg1)
66 ///     %1 = SelectOpTy(%0, %arg0, %arg1)  // %arg0, %arg1 may be swapped here.
67 ///     scf.reduce.return %1
68 template <
69     typename CompareOpTy, typename SelectOpTy,
70     typename Predicate = decltype(std::declval<CompareOpTy>().getPredicate())>
71 static bool
72 matchSelectReduction(Block &block, ArrayRef<Predicate> lessThanPredicates,
73                      ArrayRef<Predicate> greaterThanPredicates, bool &isMin) {
74   static_assert(
75       llvm::is_one_of<SelectOpTy, arith::SelectOp, LLVM::SelectOp>::value,
76       "only arithmetic and llvm select ops are supported");
77 
78   // Expect exactly three operations in the block.
79   if (block.empty() || llvm::hasSingleElement(block) ||
80       std::next(block.begin(), 2) == block.end() ||
81       std::next(block.begin(), 3) != block.end())
82     return false;
83 
84   // Check op kinds.
85   auto compare = dyn_cast<CompareOpTy>(block.front());
86   auto select = dyn_cast<SelectOpTy>(block.front().getNextNode());
87   auto terminator = dyn_cast<scf::ReduceReturnOp>(block.back());
88   if (!compare || !select || !terminator)
89     return false;
90 
91   // Block arguments must be compared.
92   if (compare->getOperands() != block.getArguments())
93     return false;
94 
95   // Detect whether the comparison is less-than or greater-than, otherwise bail.
96   bool isLess;
97   if (llvm::is_contained(lessThanPredicates, compare.getPredicate())) {
98     isLess = true;
99   } else if (llvm::is_contained(greaterThanPredicates,
100                                 compare.getPredicate())) {
101     isLess = false;
102   } else {
103     return false;
104   }
105 
106   if (select.getCondition() != compare.getResult())
107     return false;
108 
109   // Detect if the operands are swapped between cmpf and select. Match the
110   // comparison type with the requested type or with the opposite of the
111   // requested type if the operands are swapped. Use generic accessors because
112   // std and LLVM versions of select have different operand names but identical
113   // positions.
114   constexpr unsigned kTrueValue = 1;
115   constexpr unsigned kFalseValue = 2;
116   bool sameOperands = select.getOperand(kTrueValue) == compare.getLhs() &&
117                       select.getOperand(kFalseValue) == compare.getRhs();
118   bool swappedOperands = select.getOperand(kTrueValue) == compare.getRhs() &&
119                          select.getOperand(kFalseValue) == compare.getLhs();
120   if (!sameOperands && !swappedOperands)
121     return false;
122 
123   if (select.getResult() != terminator.getResult())
124     return false;
125 
126   // The reduction is a min if it uses less-than predicates with same operands
127   // or greather-than predicates with swapped operands. Similarly for max.
128   isMin = (isLess && sameOperands) || (!isLess && swappedOperands);
129   return isMin || (isLess & swappedOperands) || (!isLess && sameOperands);
130 }
131 
132 /// Returns the float semantics for the given float type.
133 static const llvm::fltSemantics &fltSemanticsForType(FloatType type) {
134   if (type.isF16())
135     return llvm::APFloat::IEEEhalf();
136   if (type.isF32())
137     return llvm::APFloat::IEEEsingle();
138   if (type.isF64())
139     return llvm::APFloat::IEEEdouble();
140   if (type.isF128())
141     return llvm::APFloat::IEEEquad();
142   if (type.isBF16())
143     return llvm::APFloat::BFloat();
144   if (type.isF80())
145     return llvm::APFloat::x87DoubleExtended();
146   llvm_unreachable("unknown float type");
147 }
148 
149 /// Returns an attribute with the minimum (if `min` is set) or the maximum value
150 /// (otherwise) for the given float type.
151 static Attribute minMaxValueForFloat(Type type, bool min) {
152   auto fltType = type.cast<FloatType>();
153   return FloatAttr::get(
154       type, llvm::APFloat::getLargest(fltSemanticsForType(fltType), min));
155 }
156 
157 /// Returns an attribute with the signed integer minimum (if `min` is set) or
158 /// the maximum value (otherwise) for the given integer type, regardless of its
159 /// signedness semantics (only the width is considered).
160 static Attribute minMaxValueForSignedInt(Type type, bool min) {
161   auto intType = type.cast<IntegerType>();
162   unsigned bitwidth = intType.getWidth();
163   return IntegerAttr::get(type, min ? llvm::APInt::getSignedMinValue(bitwidth)
164                                     : llvm::APInt::getSignedMaxValue(bitwidth));
165 }
166 
167 /// Returns an attribute with the unsigned integer minimum (if `min` is set) or
168 /// the maximum value (otherwise) for the given integer type, regardless of its
169 /// signedness semantics (only the width is considered).
170 static Attribute minMaxValueForUnsignedInt(Type type, bool min) {
171   auto intType = type.cast<IntegerType>();
172   unsigned bitwidth = intType.getWidth();
173   return IntegerAttr::get(type, min ? llvm::APInt::getNullValue(bitwidth)
174                                     : llvm::APInt::getAllOnesValue(bitwidth));
175 }
176 
177 /// Creates an OpenMP reduction declaration and inserts it into the provided
178 /// symbol table. The declaration has a constant initializer with the neutral
179 /// value `initValue`, and the reduction combiner carried over from `reduce`.
180 static omp::ReductionDeclareOp createDecl(PatternRewriter &builder,
181                                           SymbolTable &symbolTable,
182                                           scf::ReduceOp reduce,
183                                           Attribute initValue) {
184   OpBuilder::InsertionGuard guard(builder);
185   auto decl = builder.create<omp::ReductionDeclareOp>(
186       reduce.getLoc(), "__scf_reduction", reduce.getOperand().getType());
187   symbolTable.insert(decl);
188 
189   Type type = reduce.getOperand().getType();
190   builder.createBlock(&decl.initializerRegion(), decl.initializerRegion().end(),
191                       {type}, {reduce.getOperand().getLoc()});
192   builder.setInsertionPointToEnd(&decl.initializerRegion().back());
193   Value init =
194       builder.create<LLVM::ConstantOp>(reduce.getLoc(), type, initValue);
195   builder.create<omp::YieldOp>(reduce.getLoc(), init);
196 
197   Operation *terminator = &reduce.getRegion().front().back();
198   assert(isa<scf::ReduceReturnOp>(terminator) &&
199          "expected reduce op to be terminated by redure return");
200   builder.setInsertionPoint(terminator);
201   builder.replaceOpWithNewOp<omp::YieldOp>(terminator,
202                                            terminator->getOperands());
203   builder.inlineRegionBefore(reduce.getRegion(), decl.reductionRegion(),
204                              decl.reductionRegion().end());
205   return decl;
206 }
207 
208 /// Adds an atomic reduction combiner to the given OpenMP reduction declaration
209 /// using llvm.atomicrmw of the given kind.
210 static omp::ReductionDeclareOp addAtomicRMW(OpBuilder &builder,
211                                             LLVM::AtomicBinOp atomicKind,
212                                             omp::ReductionDeclareOp decl,
213                                             scf::ReduceOp reduce) {
214   OpBuilder::InsertionGuard guard(builder);
215   Type type = reduce.getOperand().getType();
216   Type ptrType = LLVM::LLVMPointerType::get(type);
217   Location reduceOperandLoc = reduce.getOperand().getLoc();
218   builder.createBlock(&decl.atomicReductionRegion(),
219                       decl.atomicReductionRegion().end(), {ptrType, ptrType},
220                       {reduceOperandLoc, reduceOperandLoc});
221   Block *atomicBlock = &decl.atomicReductionRegion().back();
222   builder.setInsertionPointToEnd(atomicBlock);
223   Value loaded = builder.create<LLVM::LoadOp>(reduce.getLoc(),
224                                               atomicBlock->getArgument(1));
225   builder.create<LLVM::AtomicRMWOp>(reduce.getLoc(), type, atomicKind,
226                                     atomicBlock->getArgument(0), loaded,
227                                     LLVM::AtomicOrdering::monotonic);
228   builder.create<omp::YieldOp>(reduce.getLoc(), ArrayRef<Value>());
229   return decl;
230 }
231 
232 /// Creates an OpenMP reduction declaration that corresponds to the given SCF
233 /// reduction and returns it. Recognizes common reductions in order to identify
234 /// the neutral value, necessary for the OpenMP declaration. If the reduction
235 /// cannot be recognized, returns null.
236 static omp::ReductionDeclareOp declareReduction(PatternRewriter &builder,
237                                                 scf::ReduceOp reduce) {
238   Operation *container = SymbolTable::getNearestSymbolTable(reduce);
239   SymbolTable symbolTable(container);
240 
241   // Insert reduction declarations in the symbol-table ancestor before the
242   // ancestor of the current insertion point.
243   Operation *insertionPoint = reduce;
244   while (insertionPoint->getParentOp() != container)
245     insertionPoint = insertionPoint->getParentOp();
246   OpBuilder::InsertionGuard guard(builder);
247   builder.setInsertionPoint(insertionPoint);
248 
249   assert(llvm::hasSingleElement(reduce.getRegion()) &&
250          "expected reduction region to have a single element");
251 
252   // Match simple binary reductions that can be expressed with atomicrmw.
253   Type type = reduce.getOperand().getType();
254   Block &reduction = reduce.getRegion().front();
255   if (matchSimpleReduction<arith::AddFOp, LLVM::FAddOp>(reduction)) {
256     omp::ReductionDeclareOp decl = createDecl(builder, symbolTable, reduce,
257                                               builder.getFloatAttr(type, 0.0));
258     return addAtomicRMW(builder, LLVM::AtomicBinOp::fadd, decl, reduce);
259   }
260   if (matchSimpleReduction<arith::AddIOp, LLVM::AddOp>(reduction)) {
261     omp::ReductionDeclareOp decl = createDecl(builder, symbolTable, reduce,
262                                               builder.getIntegerAttr(type, 0));
263     return addAtomicRMW(builder, LLVM::AtomicBinOp::add, decl, reduce);
264   }
265   if (matchSimpleReduction<arith::OrIOp, LLVM::OrOp>(reduction)) {
266     omp::ReductionDeclareOp decl = createDecl(builder, symbolTable, reduce,
267                                               builder.getIntegerAttr(type, 0));
268     return addAtomicRMW(builder, LLVM::AtomicBinOp::_or, decl, reduce);
269   }
270   if (matchSimpleReduction<arith::XOrIOp, LLVM::XOrOp>(reduction)) {
271     omp::ReductionDeclareOp decl = createDecl(builder, symbolTable, reduce,
272                                               builder.getIntegerAttr(type, 0));
273     return addAtomicRMW(builder, LLVM::AtomicBinOp::_xor, decl, reduce);
274   }
275   if (matchSimpleReduction<arith::AndIOp, LLVM::AndOp>(reduction)) {
276     omp::ReductionDeclareOp decl = createDecl(
277         builder, symbolTable, reduce,
278         builder.getIntegerAttr(
279             type, llvm::APInt::getAllOnesValue(type.getIntOrFloatBitWidth())));
280     return addAtomicRMW(builder, LLVM::AtomicBinOp::_and, decl, reduce);
281   }
282 
283   // Match simple binary reductions that cannot be expressed with atomicrmw.
284   // TODO: add atomic region using cmpxchg (which needs atomic load to be
285   // available as an op).
286   if (matchSimpleReduction<arith::MulFOp, LLVM::FMulOp>(reduction)) {
287     return createDecl(builder, symbolTable, reduce,
288                       builder.getFloatAttr(type, 1.0));
289   }
290 
291   // Match select-based min/max reductions.
292   bool isMin;
293   if (matchSelectReduction<arith::CmpFOp, arith::SelectOp>(
294           reduction, {arith::CmpFPredicate::OLT, arith::CmpFPredicate::OLE},
295           {arith::CmpFPredicate::OGT, arith::CmpFPredicate::OGE}, isMin) ||
296       matchSelectReduction<LLVM::FCmpOp, LLVM::SelectOp>(
297           reduction, {LLVM::FCmpPredicate::olt, LLVM::FCmpPredicate::ole},
298           {LLVM::FCmpPredicate::ogt, LLVM::FCmpPredicate::oge}, isMin)) {
299     return createDecl(builder, symbolTable, reduce,
300                       minMaxValueForFloat(type, !isMin));
301   }
302   if (matchSelectReduction<arith::CmpIOp, arith::SelectOp>(
303           reduction, {arith::CmpIPredicate::slt, arith::CmpIPredicate::sle},
304           {arith::CmpIPredicate::sgt, arith::CmpIPredicate::sge}, isMin) ||
305       matchSelectReduction<LLVM::ICmpOp, LLVM::SelectOp>(
306           reduction, {LLVM::ICmpPredicate::slt, LLVM::ICmpPredicate::sle},
307           {LLVM::ICmpPredicate::sgt, LLVM::ICmpPredicate::sge}, isMin)) {
308     omp::ReductionDeclareOp decl = createDecl(
309         builder, symbolTable, reduce, minMaxValueForSignedInt(type, !isMin));
310     return addAtomicRMW(builder,
311                         isMin ? LLVM::AtomicBinOp::min : LLVM::AtomicBinOp::max,
312                         decl, reduce);
313   }
314   if (matchSelectReduction<arith::CmpIOp, arith::SelectOp>(
315           reduction, {arith::CmpIPredicate::ult, arith::CmpIPredicate::ule},
316           {arith::CmpIPredicate::ugt, arith::CmpIPredicate::uge}, isMin) ||
317       matchSelectReduction<LLVM::ICmpOp, LLVM::SelectOp>(
318           reduction, {LLVM::ICmpPredicate::ugt, LLVM::ICmpPredicate::ule},
319           {LLVM::ICmpPredicate::ugt, LLVM::ICmpPredicate::uge}, isMin)) {
320     omp::ReductionDeclareOp decl = createDecl(
321         builder, symbolTable, reduce, minMaxValueForUnsignedInt(type, !isMin));
322     return addAtomicRMW(
323         builder, isMin ? LLVM::AtomicBinOp::umin : LLVM::AtomicBinOp::umax,
324         decl, reduce);
325   }
326 
327   return nullptr;
328 }
329 
330 namespace {
331 
332 struct ParallelOpLowering : public OpRewritePattern<scf::ParallelOp> {
333   using OpRewritePattern<scf::ParallelOp>::OpRewritePattern;
334 
335   LogicalResult matchAndRewrite(scf::ParallelOp parallelOp,
336                                 PatternRewriter &rewriter) const override {
337     // Declare reductions.
338     // TODO: consider checking it here is already a compatible reduction
339     // declaration and use it instead of redeclaring.
340     SmallVector<Attribute> reductionDeclSymbols;
341     for (auto reduce : parallelOp.getOps<scf::ReduceOp>()) {
342       omp::ReductionDeclareOp decl = declareReduction(rewriter, reduce);
343       if (!decl)
344         return failure();
345       reductionDeclSymbols.push_back(
346           SymbolRefAttr::get(rewriter.getContext(), decl.sym_name()));
347     }
348 
349     // Allocate reduction variables. Make sure the we don't overflow the stack
350     // with local `alloca`s by saving and restoring the stack pointer.
351     Location loc = parallelOp.getLoc();
352     Value one = rewriter.create<LLVM::ConstantOp>(
353         loc, rewriter.getIntegerType(64), rewriter.getI64IntegerAttr(1));
354     SmallVector<Value> reductionVariables;
355     reductionVariables.reserve(parallelOp.getNumReductions());
356     for (Value init : parallelOp.getInitVals()) {
357       assert((LLVM::isCompatibleType(init.getType()) ||
358               init.getType().isa<LLVM::PointerElementTypeInterface>()) &&
359              "cannot create a reduction variable if the type is not an LLVM "
360              "pointer element");
361       Value storage = rewriter.create<LLVM::AllocaOp>(
362           loc, LLVM::LLVMPointerType::get(init.getType()), one, 0);
363       rewriter.create<LLVM::StoreOp>(loc, init, storage);
364       reductionVariables.push_back(storage);
365     }
366 
367     // Replace the reduction operations contained in this loop. Must be done
368     // here rather than in a separate pattern to have access to the list of
369     // reduction variables.
370     for (auto pair :
371          llvm::zip(parallelOp.getOps<scf::ReduceOp>(), reductionVariables)) {
372       OpBuilder::InsertionGuard guard(rewriter);
373       scf::ReduceOp reduceOp = std::get<0>(pair);
374       rewriter.setInsertionPoint(reduceOp);
375       rewriter.replaceOpWithNewOp<omp::ReductionOp>(
376           reduceOp, reduceOp.getOperand(), std::get<1>(pair));
377     }
378 
379     // Create the parallel wrapper.
380     auto ompParallel = rewriter.create<omp::ParallelOp>(loc);
381     {
382 
383       OpBuilder::InsertionGuard guard(rewriter);
384       rewriter.createBlock(&ompParallel.region());
385 
386       // Replace the loop.
387       {
388         OpBuilder::InsertionGuard allocaGuard(rewriter);
389         auto loop = rewriter.create<omp::WsLoopOp>(
390             parallelOp.getLoc(), parallelOp.getLowerBound(),
391             parallelOp.getUpperBound(), parallelOp.getStep());
392         rewriter.create<omp::TerminatorOp>(loc);
393 
394         rewriter.inlineRegionBefore(parallelOp.getRegion(), loop.region(),
395                                     loop.region().begin());
396 
397         Block *ops = rewriter.splitBlock(&*loop.region().begin(),
398                                          loop.region().begin()->begin());
399 
400         rewriter.setInsertionPointToStart(&*loop.region().begin());
401 
402         auto scope = rewriter.create<memref::AllocaScopeOp>(parallelOp.getLoc(),
403                                                             TypeRange());
404         rewriter.create<omp::YieldOp>(loc, ValueRange());
405         Block *scopeBlock = rewriter.createBlock(&scope.getBodyRegion());
406         rewriter.mergeBlocks(ops, scopeBlock);
407         auto oldYield = cast<scf::YieldOp>(scopeBlock->getTerminator());
408         rewriter.setInsertionPointToEnd(&*scope.getBodyRegion().begin());
409         rewriter.replaceOpWithNewOp<memref::AllocaScopeReturnOp>(
410             oldYield, oldYield->getOperands());
411         if (!reductionVariables.empty()) {
412           loop.reductionsAttr(
413               ArrayAttr::get(rewriter.getContext(), reductionDeclSymbols));
414           loop.reduction_varsMutable().append(reductionVariables);
415         }
416       }
417     }
418 
419     // Load loop results.
420     SmallVector<Value> results;
421     results.reserve(reductionVariables.size());
422     for (Value variable : reductionVariables) {
423       Value res = rewriter.create<LLVM::LoadOp>(loc, variable);
424       results.push_back(res);
425     }
426     rewriter.replaceOp(parallelOp, results);
427 
428     return success();
429   }
430 };
431 
432 /// Applies the conversion patterns in the given function.
433 static LogicalResult applyPatterns(ModuleOp module) {
434   ConversionTarget target(*module.getContext());
435   target.addIllegalOp<scf::ReduceOp, scf::ReduceReturnOp, scf::ParallelOp>();
436   target.addLegalDialect<omp::OpenMPDialect, LLVM::LLVMDialect,
437                          memref::MemRefDialect>();
438 
439   RewritePatternSet patterns(module.getContext());
440   patterns.add<ParallelOpLowering>(module.getContext());
441   FrozenRewritePatternSet frozen(std::move(patterns));
442   return applyPartialConversion(module, target, frozen);
443 }
444 
445 /// A pass converting SCF operations to OpenMP operations.
446 struct SCFToOpenMPPass : public ConvertSCFToOpenMPBase<SCFToOpenMPPass> {
447   /// Pass entry point.
448   void runOnOperation() override {
449     if (failed(applyPatterns(getOperation())))
450       signalPassFailure();
451   }
452 };
453 
454 } // namespace
455 
456 std::unique_ptr<OperationPass<ModuleOp>> mlir::createConvertSCFToOpenMPPass() {
457   return std::make_unique<SCFToOpenMPPass>();
458 }
459