flang/lib/Optimizer/Transforms/AffinePromotion.cpp

   1 //===-- AffinePromotion.cpp -----------------------------------------------===//
   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 transformation is a prototype that promote FIR loops operations
  10 // to affine dialect operations.
  11 // It is not part of the production pipeline and would need more work in order
  12 // to be used in production.
  13 // More information can be found in this presentation:
  14 // https://slides.com/rajanwalia/deck
  15 //
  16 //===----------------------------------------------------------------------===//
  17
  18 #include "flang/Optimizer/Dialect/FIRDialect.h"
  19 #include "flang/Optimizer/Dialect/FIROps.h"
  20 #include "flang/Optimizer/Dialect/FIRType.h"
  21 #include "flang/Optimizer/Transforms/Passes.h"
  22 #include "mlir/Dialect/Affine/IR/AffineOps.h"
  23 #include "mlir/Dialect/Func/IR/FuncOps.h"
  24 #include "mlir/Dialect/SCF/IR/SCF.h"
  25 #include "mlir/IR/BuiltinAttributes.h"
  26 #include "mlir/IR/IntegerSet.h"
  27 #include "mlir/IR/Visitors.h"
  28 #include "mlir/Transforms/DialectConversion.h"
  29 #include "llvm/ADT/DenseMap.h"
  30 #include "llvm/Support/Debug.h"
  31 #include <optional>
  32
  33 namespace fir {
  34 #define GEN_PASS_DEF_AFFINEDIALECTPROMOTION
  35 #include "flang/Optimizer/Transforms/Passes.h.inc"
  36 } // namespace fir
  37
  38 #define DEBUG_TYPE "flang-affine-promotion"
  39
  40 using namespace fir;
  41 using namespace mlir;
  42
  43 namespace {
  44 struct AffineLoopAnalysis;
  45 struct AffineIfAnalysis;
  46
  47 /// Stores analysis objects for all loops and if operations inside a function
  48 /// these analysis are used twice, first for marking operations for rewrite and
  49 /// second when doing rewrite.
  50 struct AffineFunctionAnalysis {
  51   explicit AffineFunctionAnalysis(mlir::func::FuncOp funcOp) {
  52     for (fir::DoLoopOp op : funcOp.getOps<fir::DoLoopOp>())
  53       loopAnalysisMap.try_emplace(op, op, *this);
  54   }
  55
  56   AffineLoopAnalysis getChildLoopAnalysis(fir::DoLoopOp op) const;
  57
  58   AffineIfAnalysis getChildIfAnalysis(fir::IfOp op) const;
  59
  60   llvm::DenseMap<mlir::Operation *, AffineLoopAnalysis> loopAnalysisMap;
  61   llvm::DenseMap<mlir::Operation *, AffineIfAnalysis> ifAnalysisMap;
  62 };
  63 } // namespace
  64
  65 static bool analyzeCoordinate(mlir::Value coordinate, mlir::Operation *op) {
  66   if (auto blockArg = coordinate.dyn_cast<mlir::BlockArgument>()) {
  67     if (isa<fir::DoLoopOp>(blockArg.getOwner()->getParentOp()))
  68       return true;
  69     LLVM_DEBUG(llvm::dbgs() << "AffineLoopAnalysis: array coordinate is not a "
  70                                "loop induction variable (owner not loopOp)\n";
  71                op->dump());
  72     return false;
  73   }
  74   LLVM_DEBUG(
  75       llvm::dbgs() << "AffineLoopAnalysis: array coordinate is not a loop "
  76                       "induction variable (not a block argument)\n";
  77       op->dump(); coordinate.getDefiningOp()->dump());
  78   return false;
  79 }
  80
  81 namespace {
  82 struct AffineLoopAnalysis {
  83   AffineLoopAnalysis() = default;
  84
  85   explicit AffineLoopAnalysis(fir::DoLoopOp op, AffineFunctionAnalysis &afa)
  86       : legality(analyzeLoop(op, afa)) {}
  87
  88   bool canPromoteToAffine() { return legality; }
  89
  90 private:
  91   bool analyzeBody(fir::DoLoopOp loopOperation,
  92                    AffineFunctionAnalysis &functionAnalysis) {
  93     for (auto loopOp : loopOperation.getOps<fir::DoLoopOp>()) {
  94       auto analysis = functionAnalysis.loopAnalysisMap
  95                           .try_emplace(loopOp, loopOp, functionAnalysis)
  96                           .first->getSecond();
  97       if (!analysis.canPromoteToAffine())
  98         return false;
  99     }
 100     for (auto ifOp : loopOperation.getOps<fir::IfOp>())
 101       functionAnalysis.ifAnalysisMap.try_emplace(ifOp, ifOp, functionAnalysis);
 102     return true;
 103   }
 104
 105   bool analyzeLoop(fir::DoLoopOp loopOperation,
 106                    AffineFunctionAnalysis &functionAnalysis) {
 107     LLVM_DEBUG(llvm::dbgs() << "AffineLoopAnalysis: \n"; loopOperation.dump(););
 108     return analyzeMemoryAccess(loopOperation) &&
 109            analyzeBody(loopOperation, functionAnalysis);
 110   }
 111
 112   bool analyzeReference(mlir::Value memref, mlir::Operation *op) {
 113     if (auto acoOp = memref.getDefiningOp<ArrayCoorOp>()) {
 114       if (acoOp.getMemref().getType().isa<fir::BoxType>()) {
 115         // TODO: Look if and how fir.box can be promoted to affine.
 116         LLVM_DEBUG(llvm::dbgs() << "AffineLoopAnalysis: cannot promote loop, "
 117                                    "array memory operation uses fir.box\n";
 118                    op->dump(); acoOp.dump(););
 119         return false;
 120       }
 121       bool canPromote = true;
 122       for (auto coordinate : acoOp.getIndices())
 123         canPromote = canPromote && analyzeCoordinate(coordinate, op);
 124       return canPromote;
 125     }
 126     if (auto coOp = memref.getDefiningOp<CoordinateOp>()) {
 127       LLVM_DEBUG(llvm::dbgs()
 128                      << "AffineLoopAnalysis: cannot promote loop, "
 129                         "array memory operation uses non ArrayCoorOp\n";
 130                  op->dump(); coOp.dump(););
 131
 132       return false;
 133     }
 134     LLVM_DEBUG(llvm::dbgs() << "AffineLoopAnalysis: unknown type of memory "
 135                                "reference for array load\n";
 136                op->dump(););
 137     return false;
 138   }
 139
 140   bool analyzeMemoryAccess(fir::DoLoopOp loopOperation) {
 141     for (auto loadOp : loopOperation.getOps<fir::LoadOp>())
 142       if (!analyzeReference(loadOp.getMemref(), loadOp))
 143         return false;
 144     for (auto storeOp : loopOperation.getOps<fir::StoreOp>())
 145       if (!analyzeReference(storeOp.getMemref(), storeOp))
 146         return false;
 147     return true;
 148   }
 149
 150   bool legality{};
 151 };
 152 } // namespace
 153
 154 AffineLoopAnalysis
 155 AffineFunctionAnalysis::getChildLoopAnalysis(fir::DoLoopOp op) const {
 156   auto it = loopAnalysisMap.find_as(op);
 157   if (it == loopAnalysisMap.end()) {
 158     LLVM_DEBUG(llvm::dbgs() << "AffineFunctionAnalysis: not computed for:\n";
 159                op.dump(););
 160     op.emitError("error in fetching loop analysis in AffineFunctionAnalysis\n");
 161     return {};
 162   }
 163   return it->getSecond();
 164 }
 165
 166 namespace {
 167 /// Calculates arguments for creating an IntegerSet. symCount, dimCount are the
 168 /// final number of symbols and dimensions of the affine map. Integer set if
 169 /// possible is in Optional IntegerSet.
 170 struct AffineIfCondition {
 171   using MaybeAffineExpr = std::optional<mlir::AffineExpr>;
 172
 173   explicit AffineIfCondition(mlir::Value fc) : firCondition(fc) {
 174     if (auto condDef = firCondition.getDefiningOp<mlir::arith::CmpIOp>())
 175       fromCmpIOp(condDef);
 176   }
 177
 178   bool hasIntegerSet() const { return integerSet.has_value(); }
 179
 180   mlir::IntegerSet getIntegerSet() const {
 181     assert(hasIntegerSet() && "integer set is missing");
 182     return *integerSet;
 183   }
 184
 185   mlir::ValueRange getAffineArgs() const { return affineArgs; }
 186
 187 private:
 188   MaybeAffineExpr affineBinaryOp(mlir::AffineExprKind kind, mlir::Value lhs,
 189                                  mlir::Value rhs) {
 190     return affineBinaryOp(kind, toAffineExpr(lhs), toAffineExpr(rhs));
 191   }
 192
 193   MaybeAffineExpr affineBinaryOp(mlir::AffineExprKind kind, MaybeAffineExpr lhs,
 194                                  MaybeAffineExpr rhs) {
 195     if (lhs && rhs)
 196       return mlir::getAffineBinaryOpExpr(kind, *lhs, *rhs);
 197     return {};
 198   }
 199
 200   MaybeAffineExpr toAffineExpr(MaybeAffineExpr e) { return e; }
 201
 202   MaybeAffineExpr toAffineExpr(int64_t value) {
 203     return {mlir::getAffineConstantExpr(value, firCondition.getContext())};
 204   }
 205
 206   /// Returns an AffineExpr if it is a result of operations that can be done
 207   /// in an affine expression, this includes -, +, *, rem, constant.
 208   /// block arguments of a loopOp or forOp are used as dimensions
 209   MaybeAffineExpr toAffineExpr(mlir::Value value) {
 210     if (auto op = value.getDefiningOp<mlir::arith::SubIOp>())
 211       return affineBinaryOp(
 212           mlir::AffineExprKind::Add, toAffineExpr(op.getLhs()),
 213           affineBinaryOp(mlir::AffineExprKind::Mul, toAffineExpr(op.getRhs()),
 214                          toAffineExpr(-1)));
 215     if (auto op = value.getDefiningOp<mlir::arith::AddIOp>())
 216       return affineBinaryOp(mlir::AffineExprKind::Add, op.getLhs(),
 217                             op.getRhs());
 218     if (auto op = value.getDefiningOp<mlir::arith::MulIOp>())
 219       return affineBinaryOp(mlir::AffineExprKind::Mul, op.getLhs(),
 220                             op.getRhs());
 221     if (auto op = value.getDefiningOp<mlir::arith::RemUIOp>())
 222       return affineBinaryOp(mlir::AffineExprKind::Mod, op.getLhs(),
 223                             op.getRhs());
 224     if (auto op = value.getDefiningOp<mlir::arith::ConstantOp>())
 225       if (auto intConstant = op.getValue().dyn_cast<IntegerAttr>())
 226         return toAffineExpr(intConstant.getInt());
 227     if (auto blockArg = value.dyn_cast<mlir::BlockArgument>()) {
 228       affineArgs.push_back(value);
 229       if (isa<fir::DoLoopOp>(blockArg.getOwner()->getParentOp()) ||
 230           isa<mlir::AffineForOp>(blockArg.getOwner()->getParentOp()))
 231         return {mlir::getAffineDimExpr(dimCount++, value.getContext())};
 232       return {mlir::getAffineSymbolExpr(symCount++, value.getContext())};
 233     }
 234     return {};
 235   }
 236
 237   void fromCmpIOp(mlir::arith::CmpIOp cmpOp) {
 238     auto lhsAffine = toAffineExpr(cmpOp.getLhs());
 239     auto rhsAffine = toAffineExpr(cmpOp.getRhs());
 240     if (!lhsAffine || !rhsAffine)
 241       return;
 242     auto constraintPair =
 243         constraint(cmpOp.getPredicate(), *rhsAffine - *lhsAffine);
 244     if (!constraintPair)
 245       return;
 246     integerSet = mlir::IntegerSet::get(
 247         dimCount, symCount, {constraintPair->first}, {constraintPair->second});
 248   }
 249
 250   std::optional<std::pair<AffineExpr, bool>>
 251   constraint(mlir::arith::CmpIPredicate predicate, mlir::AffineExpr basic) {
 252     switch (predicate) {
 253     case mlir::arith::CmpIPredicate::slt:
 254       return {std::make_pair(basic - 1, false)};
 255     case mlir::arith::CmpIPredicate::sle:
 256       return {std::make_pair(basic, false)};
 257     case mlir::arith::CmpIPredicate::sgt:
 258       return {std::make_pair(1 - basic, false)};
 259     case mlir::arith::CmpIPredicate::sge:
 260       return {std::make_pair(0 - basic, false)};
 261     case mlir::arith::CmpIPredicate::eq:
 262       return {std::make_pair(basic, true)};
 263     default:
 264       return {};
 265     }
 266   }
 267
 268   llvm::SmallVector<mlir::Value> affineArgs;
 269   std::optional<mlir::IntegerSet> integerSet;
 270   mlir::Value firCondition;
 271   unsigned symCount{0u};
 272   unsigned dimCount{0u};
 273 };
 274 } // namespace
 275
 276 namespace {
 277 /// Analysis for affine promotion of fir.if
 278 struct AffineIfAnalysis {
 279   AffineIfAnalysis() = default;
 280
 281   explicit AffineIfAnalysis(fir::IfOp op, AffineFunctionAnalysis &afa)
 282       : legality(analyzeIf(op, afa)) {}
 283
 284   bool canPromoteToAffine() { return legality; }
 285
 286 private:
 287   bool analyzeIf(fir::IfOp op, AffineFunctionAnalysis &afa) {
 288     if (op.getNumResults() == 0)
 289       return true;
 290     LLVM_DEBUG(llvm::dbgs()
 291                    << "AffineIfAnalysis: not promoting as op has results\n";);
 292     return false;
 293   }
 294
 295   bool legality{};
 296 };
 297 } // namespace
 298
 299 AffineIfAnalysis
 300 AffineFunctionAnalysis::getChildIfAnalysis(fir::IfOp op) const {
 301   auto it = ifAnalysisMap.find_as(op);
 302   if (it == ifAnalysisMap.end()) {
 303     LLVM_DEBUG(llvm::dbgs() << "AffineFunctionAnalysis: not computed for:\n";
 304                op.dump(););
 305     op.emitError("error in fetching if analysis in AffineFunctionAnalysis\n");
 306     return {};
 307   }
 308   return it->getSecond();
 309 }
 310
 311 /// AffineMap rewriting fir.array_coor operation to affine apply,
 312 /// %dim = fir.gendim %lowerBound, %upperBound, %stride
 313 /// %a = fir.array_coor %arr(%dim) %i
 314 /// returning affineMap = affine_map<(i)[lb, ub, st] -> (i*st - lb)>
 315 static mlir::AffineMap createArrayIndexAffineMap(unsigned dimensions,
 316                                                  MLIRContext *context) {
 317   auto index = mlir::getAffineConstantExpr(0, context);
 318   auto accuExtent = mlir::getAffineConstantExpr(1, context);
 319   for (unsigned i = 0; i < dimensions; ++i) {
 320     mlir::AffineExpr idx = mlir::getAffineDimExpr(i, context),
 321                      lowerBound = mlir::getAffineSymbolExpr(i * 3, context),
 322                      currentExtent =
 323                          mlir::getAffineSymbolExpr(i * 3 + 1, context),
 324                      stride = mlir::getAffineSymbolExpr(i * 3 + 2, context),
 325                      currentPart = (idx * stride - lowerBound) * accuExtent;
 326     index = currentPart + index;
 327     accuExtent = accuExtent * currentExtent;
 328   }
 329   return mlir::AffineMap::get(dimensions, dimensions * 3, index);
 330 }
 331
 332 static std::optional<int64_t> constantIntegerLike(const mlir::Value value) {
 333   if (auto definition = value.getDefiningOp<mlir::arith::ConstantOp>())
 334     if (auto stepAttr = definition.getValue().dyn_cast<IntegerAttr>())
 335       return stepAttr.getInt();
 336   return {};
 337 }
 338
 339 static mlir::Type coordinateArrayElement(fir::ArrayCoorOp op) {
 340   if (auto refType =
 341           op.getMemref().getType().dyn_cast_or_null<ReferenceType>()) {
 342     if (auto seqType = refType.getEleTy().dyn_cast_or_null<SequenceType>()) {
 343       return seqType.getEleTy();
 344     }
 345   }
 346   op.emitError(
 347       "AffineLoopConversion: array type in coordinate operation not valid\n");
 348   return mlir::Type();
 349 }
 350
 351 static void populateIndexArgs(fir::ArrayCoorOp acoOp, fir::ShapeOp shape,
 352                               SmallVectorImpl<mlir::Value> &indexArgs,
 353                               mlir::PatternRewriter &rewriter) {
 354   auto one = rewriter.create<mlir::arith::ConstantOp>(
 355       acoOp.getLoc(), rewriter.getIndexType(), rewriter.getIndexAttr(1));
 356   auto extents = shape.getExtents();
 357   for (auto i = extents.begin(); i < extents.end(); i++) {
 358     indexArgs.push_back(one);
 359     indexArgs.push_back(*i);
 360     indexArgs.push_back(one);
 361   }
 362 }
 363
 364 static void populateIndexArgs(fir::ArrayCoorOp acoOp, fir::ShapeShiftOp shape,
 365                               SmallVectorImpl<mlir::Value> &indexArgs,
 366                               mlir::PatternRewriter &rewriter) {
 367   auto one = rewriter.create<mlir::arith::ConstantOp>(
 368       acoOp.getLoc(), rewriter.getIndexType(), rewriter.getIndexAttr(1));
 369   auto extents = shape.getPairs();
 370   for (auto i = extents.begin(); i < extents.end();) {
 371     indexArgs.push_back(*i++);
 372     indexArgs.push_back(*i++);
 373     indexArgs.push_back(one);
 374   }
 375 }
 376
 377 static void populateIndexArgs(fir::ArrayCoorOp acoOp, fir::SliceOp slice,
 378                               SmallVectorImpl<mlir::Value> &indexArgs,
 379                               mlir::PatternRewriter &rewriter) {
 380   auto extents = slice.getTriples();
 381   for (auto i = extents.begin(); i < extents.end();) {
 382     indexArgs.push_back(*i++);
 383     indexArgs.push_back(*i++);
 384     indexArgs.push_back(*i++);
 385   }
 386 }
 387
 388 static void populateIndexArgs(fir::ArrayCoorOp acoOp,
 389                               SmallVectorImpl<mlir::Value> &indexArgs,
 390                               mlir::PatternRewriter &rewriter) {
 391   if (auto shape = acoOp.getShape().getDefiningOp<ShapeOp>())
 392     return populateIndexArgs(acoOp, shape, indexArgs, rewriter);
 393   if (auto shapeShift = acoOp.getShape().getDefiningOp<ShapeShiftOp>())
 394     return populateIndexArgs(acoOp, shapeShift, indexArgs, rewriter);
 395   if (auto slice = acoOp.getShape().getDefiningOp<SliceOp>())
 396     return populateIndexArgs(acoOp, slice, indexArgs, rewriter);
 397 }
 398
 399 /// Returns affine.apply and fir.convert from array_coor and gendims
 400 static std::pair<mlir::AffineApplyOp, fir::ConvertOp>
 401 createAffineOps(mlir::Value arrayRef, mlir::PatternRewriter &rewriter) {
 402   auto acoOp = arrayRef.getDefiningOp<ArrayCoorOp>();
 403   auto affineMap =
 404       createArrayIndexAffineMap(acoOp.getIndices().size(), acoOp.getContext());
 405   SmallVector<mlir::Value> indexArgs;
 406   indexArgs.append(acoOp.getIndices().begin(), acoOp.getIndices().end());
 407
 408   populateIndexArgs(acoOp, indexArgs, rewriter);
 409
 410   auto affineApply = rewriter.create<mlir::AffineApplyOp>(acoOp.getLoc(),
 411                                                           affineMap, indexArgs);
 412   auto arrayElementType = coordinateArrayElement(acoOp);
 413   auto newType =
 414       mlir::MemRefType::get({mlir::ShapedType::kDynamic}, arrayElementType);
 415   auto arrayConvert = rewriter.create<fir::ConvertOp>(acoOp.getLoc(), newType,
 416                                                       acoOp.getMemref());
 417   return std::make_pair(affineApply, arrayConvert);
 418 }
 419
 420 static void rewriteLoad(fir::LoadOp loadOp, mlir::PatternRewriter &rewriter) {
 421   rewriter.setInsertionPoint(loadOp);
 422   auto affineOps = createAffineOps(loadOp.getMemref(), rewriter);
 423   rewriter.replaceOpWithNewOp<mlir::AffineLoadOp>(
 424       loadOp, affineOps.second.getResult(), affineOps.first.getResult());
 425 }
 426
 427 static void rewriteStore(fir::StoreOp storeOp,
 428                          mlir::PatternRewriter &rewriter) {
 429   rewriter.setInsertionPoint(storeOp);
 430   auto affineOps = createAffineOps(storeOp.getMemref(), rewriter);
 431   rewriter.replaceOpWithNewOp<mlir::AffineStoreOp>(storeOp, storeOp.getValue(),
 432                                                    affineOps.second.getResult(),
 433                                                    affineOps.first.getResult());
 434 }
 435
 436 static void rewriteMemoryOps(Block *block, mlir::PatternRewriter &rewriter) {
 437   for (auto &bodyOp : block->getOperations()) {
 438     if (isa<fir::LoadOp>(bodyOp))
 439       rewriteLoad(cast<fir::LoadOp>(bodyOp), rewriter);
 440     if (isa<fir::StoreOp>(bodyOp))
 441       rewriteStore(cast<fir::StoreOp>(bodyOp), rewriter);
 442   }
 443 }
 444
 445 namespace {
 446 /// Convert `fir.do_loop` to `affine.for`, creates fir.convert for arrays to
 447 /// memref, rewrites array_coor to affine.apply with affine_map. Rewrites fir
 448 /// loads and stores to affine.
 449 class AffineLoopConversion : public mlir::OpRewritePattern<fir::DoLoopOp> {
 450 public:
 451   using OpRewritePattern::OpRewritePattern;
 452   AffineLoopConversion(mlir::MLIRContext *context, AffineFunctionAnalysis &afa)
 453       : OpRewritePattern(context), functionAnalysis(afa) {}
 454
 455   mlir::LogicalResult
 456   matchAndRewrite(fir::DoLoopOp loop,
 457                   mlir::PatternRewriter &rewriter) const override {
 458     LLVM_DEBUG(llvm::dbgs() << "AffineLoopConversion: rewriting loop:\n";
 459                loop.dump(););
 460     LLVM_ATTRIBUTE_UNUSED auto loopAnalysis =
 461         functionAnalysis.getChildLoopAnalysis(loop);
 462     auto &loopOps = loop.getBody()->getOperations();
 463     auto loopAndIndex = createAffineFor(loop, rewriter);
 464     auto affineFor = loopAndIndex.first;
 465     auto inductionVar = loopAndIndex.second;
 466
 467     rewriter.startRootUpdate(affineFor.getOperation());
 468     affineFor.getBody()->getOperations().splice(
 469         std::prev(affineFor.getBody()->end()), loopOps, loopOps.begin(),
 470         std::prev(loopOps.end()));
 471     rewriter.finalizeRootUpdate(affineFor.getOperation());
 472
 473     rewriter.startRootUpdate(loop.getOperation());
 474     loop.getInductionVar().replaceAllUsesWith(inductionVar);
 475     rewriter.finalizeRootUpdate(loop.getOperation());
 476
 477     rewriteMemoryOps(affineFor.getBody(), rewriter);
 478
 479     LLVM_DEBUG(llvm::dbgs() << "AffineLoopConversion: loop rewriten to:\n";
 480                affineFor.dump(););
 481     rewriter.replaceOp(loop, affineFor.getOperation()->getResults());
 482     return success();
 483   }
 484
 485 private:
 486   std::pair<mlir::AffineForOp, mlir::Value>
 487   createAffineFor(fir::DoLoopOp op, mlir::PatternRewriter &rewriter) const {
 488     if (auto constantStep = constantIntegerLike(op.getStep()))
 489       if (*constantStep > 0)
 490         return positiveConstantStep(op, *constantStep, rewriter);
 491     return genericBounds(op, rewriter);
 492   }
 493
 494   // when step for the loop is positive compile time constant
 495   std::pair<mlir::AffineForOp, mlir::Value>
 496   positiveConstantStep(fir::DoLoopOp op, int64_t step,
 497                        mlir::PatternRewriter &rewriter) const {
 498     auto affineFor = rewriter.create<mlir::AffineForOp>(
 499         op.getLoc(), ValueRange(op.getLowerBound()),
 500         mlir::AffineMap::get(0, 1,
 501                              mlir::getAffineSymbolExpr(0, op.getContext())),
 502         ValueRange(op.getUpperBound()),
 503         mlir::AffineMap::get(0, 1,
 504                              1 + mlir::getAffineSymbolExpr(0, op.getContext())),
 505         step);
 506     return std::make_pair(affineFor, affineFor.getInductionVar());
 507   }
 508
 509   std::pair<mlir::AffineForOp, mlir::Value>
 510   genericBounds(fir::DoLoopOp op, mlir::PatternRewriter &rewriter) const {
 511     auto lowerBound = mlir::getAffineSymbolExpr(0, op.getContext());
 512     auto upperBound = mlir::getAffineSymbolExpr(1, op.getContext());
 513     auto step = mlir::getAffineSymbolExpr(2, op.getContext());
 514     mlir::AffineMap upperBoundMap = mlir::AffineMap::get(
 515         0, 3, (upperBound - lowerBound + step).floorDiv(step));
 516     auto genericUpperBound = rewriter.create<mlir::AffineApplyOp>(
 517         op.getLoc(), upperBoundMap,
 518         ValueRange({op.getLowerBound(), op.getUpperBound(), op.getStep()}));
 519     auto actualIndexMap = mlir::AffineMap::get(
 520         1, 2,
 521         (lowerBound + mlir::getAffineDimExpr(0, op.getContext())) *
 522             mlir::getAffineSymbolExpr(1, op.getContext()));
 523
 524     auto affineFor = rewriter.create<mlir::AffineForOp>(
 525         op.getLoc(), ValueRange(),
 526         AffineMap::getConstantMap(0, op.getContext()),
 527         genericUpperBound.getResult(),
 528         mlir::AffineMap::get(0, 1,
 529                              1 + mlir::getAffineSymbolExpr(0, op.getContext())),
 530         1);
 531     rewriter.setInsertionPointToStart(affineFor.getBody());
 532     auto actualIndex = rewriter.create<mlir::AffineApplyOp>(
 533         op.getLoc(), actualIndexMap,
 534         ValueRange(
 535             {affineFor.getInductionVar(), op.getLowerBound(), op.getStep()}));
 536     return std::make_pair(affineFor, actualIndex.getResult());
 537   }
 538
 539   AffineFunctionAnalysis &functionAnalysis;
 540 };
 541
 542 /// Convert `fir.if` to `affine.if`.
 543 class AffineIfConversion : public mlir::OpRewritePattern<fir::IfOp> {
 544 public:
 545   using OpRewritePattern::OpRewritePattern;
 546   AffineIfConversion(mlir::MLIRContext *context, AffineFunctionAnalysis &afa)
 547       : OpRewritePattern(context) {}
 548   mlir::LogicalResult
 549   matchAndRewrite(fir::IfOp op,
 550                   mlir::PatternRewriter &rewriter) const override {
 551     LLVM_DEBUG(llvm::dbgs() << "AffineIfConversion: rewriting if:\n";
 552                op.dump(););
 553     auto &ifOps = op.getThenRegion().front().getOperations();
 554     auto affineCondition = AffineIfCondition(op.getCondition());
 555     if (!affineCondition.hasIntegerSet()) {
 556       LLVM_DEBUG(
 557           llvm::dbgs()
 558               << "AffineIfConversion: couldn't calculate affine condition\n";);
 559       return failure();
 560     }
 561     auto affineIf = rewriter.create<mlir::AffineIfOp>(
 562         op.getLoc(), affineCondition.getIntegerSet(),
 563         affineCondition.getAffineArgs(), !op.getElseRegion().empty());
 564     rewriter.startRootUpdate(affineIf);
 565     affineIf.getThenBlock()->getOperations().splice(
 566         std::prev(affineIf.getThenBlock()->end()), ifOps, ifOps.begin(),
 567         std::prev(ifOps.end()));
 568     if (!op.getElseRegion().empty()) {
 569       auto &otherOps = op.getElseRegion().front().getOperations();
 570       affineIf.getElseBlock()->getOperations().splice(
 571           std::prev(affineIf.getElseBlock()->end()), otherOps, otherOps.begin(),
 572           std::prev(otherOps.end()));
 573     }
 574     rewriter.finalizeRootUpdate(affineIf);
 575     rewriteMemoryOps(affineIf.getBody(), rewriter);
 576
 577     LLVM_DEBUG(llvm::dbgs() << "AffineIfConversion: if converted to:\n";
 578                affineIf.dump(););
 579     rewriter.replaceOp(op, affineIf.getOperation()->getResults());
 580     return success();
 581   }
 582 };
 583
 584 /// Promote fir.do_loop and fir.if to affine.for and affine.if, in the cases
 585 /// where such a promotion is possible.
 586 class AffineDialectPromotion
 587     : public fir::impl::AffineDialectPromotionBase<AffineDialectPromotion> {
 588 public:
 589   void runOnOperation() override {
 590
 591     auto *context = &getContext();
 592     auto function = getOperation();
 593     markAllAnalysesPreserved();
 594     auto functionAnalysis = AffineFunctionAnalysis(function);
 595     mlir::RewritePatternSet patterns(context);
 596     patterns.insert<AffineIfConversion>(context, functionAnalysis);
 597     patterns.insert<AffineLoopConversion>(context, functionAnalysis);
 598     mlir::ConversionTarget target = *context;
 599     target.addLegalDialect<mlir::AffineDialect, FIROpsDialect,
 600                            mlir::scf::SCFDialect, mlir::arith::ArithDialect,
 601                            mlir::func::FuncDialect>();
 602     target.addDynamicallyLegalOp<IfOp>([&functionAnalysis](fir::IfOp op) {
 603       return !(functionAnalysis.getChildIfAnalysis(op).canPromoteToAffine());
 604     });
 605     target.addDynamicallyLegalOp<DoLoopOp>([&functionAnalysis](
 606                                                fir::DoLoopOp op) {
 607       return !(functionAnalysis.getChildLoopAnalysis(op).canPromoteToAffine());
 608     });
 609
 610     LLVM_DEBUG(llvm::dbgs()
 611                    << "AffineDialectPromotion: running promotion on: \n";
 612                function.print(llvm::dbgs()););
 613     // apply the patterns
 614     if (mlir::failed(mlir::applyPartialConversion(function, target,
 615                                                   std::move(patterns)))) {
 616       mlir::emitError(mlir::UnknownLoc::get(context),
 617                       "error in converting to affine dialect\n");
 618       signalPassFailure();
 619     }
 620   }
 621 };
 622 } // namespace
 623
 624 /// Convert FIR loop constructs to the Affine dialect
 625 std::unique_ptr<mlir::Pass> fir::createPromoteToAffinePass() {
 626   return std::make_unique<AffineDialectPromotion>();
 627 }