| 1 | //===-- AffineDemotion.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 demote affine dialects operations |
| 10 | // after optimizations to FIR loops operations. |
| 11 | // It is used after the AffinePromotion pass. |
| 12 | // It is not part of the production pipeline and would need more work in order |
| 13 | // to be used in production. |
| 14 | // More information can be found in this presentation: |
| 15 | // https://slides.com/rajanwalia/deck |
| 16 | // |
| 17 | //===----------------------------------------------------------------------===// |
| 18 | |
| 19 | #include "flang/Optimizer/Dialect/FIRDialect.h" |
| 20 | #include "flang/Optimizer/Dialect/FIROps.h" |
| 21 | #include "flang/Optimizer/Dialect/FIRType.h" |
| 22 | #include "flang/Optimizer/Transforms/Passes.h" |
| 23 | #include "mlir/Dialect/Affine/IR/AffineOps.h" |
| 24 | #include "mlir/Dialect/Affine/Utils.h" |
| 25 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
| 26 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
| 27 | #include "mlir/Dialect/SCF/IR/SCF.h" |
| 28 | #include "mlir/IR/BuiltinAttributes.h" |
| 29 | #include "mlir/IR/IntegerSet.h" |
| 30 | #include "mlir/IR/Visitors.h" |
| 31 | #include "mlir/Pass/Pass.h" |
| 32 | #include "mlir/Transforms/DialectConversion.h" |
| 33 | #include "llvm/ADT/DenseMap.h" |
| 34 | #include "llvm/Support/CommandLine.h" |
| 35 | #include "llvm/Support/Debug.h" |
| 36 | |
| 37 | namespace fir { |
| 38 | #define GEN_PASS_DEF_AFFINEDIALECTDEMOTION |
| 39 | #include "flang/Optimizer/Transforms/Passes.h.inc" |
| 40 | } // namespace fir |
| 41 | |
| 42 | #define DEBUG_TYPE "flang-affine-demotion" |
| 43 | |
| 44 | using namespace fir; |
| 45 | using namespace mlir; |
| 46 | |
| 47 | namespace { |
| 48 | |
| 49 | class AffineLoadConversion |
| 50 | : public OpConversionPattern<mlir::affine::AffineLoadOp> { |
| 51 | public: |
| 52 | using OpConversionPattern<mlir::affine::AffineLoadOp>::OpConversionPattern; |
| 53 | |
| 54 | LogicalResult |
| 55 | matchAndRewrite(mlir::affine::AffineLoadOp op, OpAdaptor adaptor, |
| 56 | ConversionPatternRewriter &rewriter) const override { |
| 57 | SmallVector<Value> indices(adaptor.getIndices()); |
| 58 | auto maybeExpandedMap = affine::expandAffineMap(rewriter, op.getLoc(), |
| 59 | op.getAffineMap(), indices); |
| 60 | if (!maybeExpandedMap) |
| 61 | return failure(); |
| 62 | |
| 63 | auto coorOp = rewriter.create<fir::CoordinateOp>( |
| 64 | op.getLoc(), fir::ReferenceType::get(op.getResult().getType()), |
| 65 | adaptor.getMemref(), *maybeExpandedMap); |
| 66 | |
| 67 | rewriter.replaceOpWithNewOp<fir::LoadOp>(op, coorOp.getResult()); |
| 68 | return success(); |
| 69 | } |
| 70 | }; |
| 71 | |
| 72 | class AffineStoreConversion |
| 73 | : public OpConversionPattern<mlir::affine::AffineStoreOp> { |
| 74 | public: |
| 75 | using OpConversionPattern<mlir::affine::AffineStoreOp>::OpConversionPattern; |
| 76 | |
| 77 | LogicalResult |
| 78 | matchAndRewrite(mlir::affine::AffineStoreOp op, OpAdaptor adaptor, |
| 79 | ConversionPatternRewriter &rewriter) const override { |
| 80 | SmallVector<Value> indices(op.getIndices()); |
| 81 | auto maybeExpandedMap = affine::expandAffineMap(rewriter, op.getLoc(), |
| 82 | op.getAffineMap(), indices); |
| 83 | if (!maybeExpandedMap) |
| 84 | return failure(); |
| 85 | |
| 86 | auto coorOp = rewriter.create<fir::CoordinateOp>( |
| 87 | op.getLoc(), fir::ReferenceType::get(op.getValueToStore().getType()), |
| 88 | adaptor.getMemref(), *maybeExpandedMap); |
| 89 | rewriter.replaceOpWithNewOp<fir::StoreOp>(op, adaptor.getValue(), |
| 90 | coorOp.getResult()); |
| 91 | return success(); |
| 92 | } |
| 93 | }; |
| 94 | |
| 95 | class ConvertConversion : public mlir::OpRewritePattern<fir::ConvertOp> { |
| 96 | public: |
| 97 | using OpRewritePattern::OpRewritePattern; |
| 98 | llvm::LogicalResult |
| 99 | matchAndRewrite(fir::ConvertOp op, |
| 100 | mlir::PatternRewriter &rewriter) const override { |
| 101 | if (mlir::isa<mlir::MemRefType>(op.getRes().getType())) { |
| 102 | // due to index calculation moving to affine maps we still need to |
| 103 | // add converts for sequence types this has a side effect of losing |
| 104 | // some information about arrays with known dimensions by creating: |
| 105 | // fir.convert %arg0 : (!fir.ref<!fir.array<5xi32>>) -> |
| 106 | // !fir.ref<!fir.array<?xi32>> |
| 107 | if (auto refTy = |
| 108 | mlir::dyn_cast<fir::ReferenceType>(op.getValue().getType())) |
| 109 | if (auto arrTy = mlir::dyn_cast<fir::SequenceType>(refTy.getEleTy())) { |
| 110 | fir::SequenceType::Shape flatShape = { |
| 111 | fir::SequenceType::getUnknownExtent()}; |
| 112 | auto flatArrTy = fir::SequenceType::get(flatShape, arrTy.getEleTy()); |
| 113 | auto flatTy = fir::ReferenceType::get(flatArrTy); |
| 114 | rewriter.replaceOpWithNewOp<fir::ConvertOp>(op, flatTy, |
| 115 | op.getValue()); |
| 116 | return success(); |
| 117 | } |
| 118 | rewriter.startOpModification(op->getParentOp()); |
| 119 | op.getResult().replaceAllUsesWith(op.getValue()); |
| 120 | rewriter.finalizeOpModification(op->getParentOp()); |
| 121 | rewriter.eraseOp(op); |
| 122 | } |
| 123 | return success(); |
| 124 | } |
| 125 | }; |
| 126 | |
| 127 | mlir::Type convertMemRef(mlir::MemRefType type) { |
| 128 | return fir::SequenceType::get(SmallVector<int64_t>(type.getShape()), |
| 129 | type.getElementType()); |
| 130 | } |
| 131 | |
| 132 | class StdAllocConversion : public mlir::OpRewritePattern<memref::AllocOp> { |
| 133 | public: |
| 134 | using OpRewritePattern::OpRewritePattern; |
| 135 | llvm::LogicalResult |
| 136 | matchAndRewrite(memref::AllocOp op, |
| 137 | mlir::PatternRewriter &rewriter) const override { |
| 138 | rewriter.replaceOpWithNewOp<fir::AllocaOp>(op, convertMemRef(op.getType()), |
| 139 | op.getMemref()); |
| 140 | return success(); |
| 141 | } |
| 142 | }; |
| 143 | |
| 144 | class AffineDialectDemotion |
| 145 | : public fir::impl::AffineDialectDemotionBase<AffineDialectDemotion> { |
| 146 | public: |
| 147 | void runOnOperation() override { |
| 148 | auto *context = &getContext(); |
| 149 | auto function = getOperation(); |
| 150 | LLVM_DEBUG(llvm::dbgs() << "AffineDemotion: running on function:\n" ; |
| 151 | function.print(llvm::dbgs());); |
| 152 | |
| 153 | mlir::RewritePatternSet patterns(context); |
| 154 | patterns.insert<ConvertConversion>(context); |
| 155 | patterns.insert<AffineLoadConversion>(context); |
| 156 | patterns.insert<AffineStoreConversion>(context); |
| 157 | patterns.insert<StdAllocConversion>(context); |
| 158 | mlir::ConversionTarget target(*context); |
| 159 | target.addIllegalOp<memref::AllocOp>(); |
| 160 | target.addDynamicallyLegalOp<fir::ConvertOp>([](fir::ConvertOp op) { |
| 161 | if (mlir::isa<mlir::MemRefType>(op.getRes().getType())) |
| 162 | return false; |
| 163 | return true; |
| 164 | }); |
| 165 | target |
| 166 | .addLegalDialect<FIROpsDialect, mlir::scf::SCFDialect, |
| 167 | mlir::arith::ArithDialect, mlir::func::FuncDialect>(); |
| 168 | |
| 169 | if (mlir::failed(mlir::applyPartialConversion(function, target, |
| 170 | std::move(patterns)))) { |
| 171 | mlir::emitError(mlir::UnknownLoc::get(context), |
| 172 | "error in converting affine dialect\n" ); |
| 173 | signalPassFailure(); |
| 174 | } |
| 175 | } |
| 176 | }; |
| 177 | |
| 178 | } // namespace |
| 179 | |
| 180 | std::unique_ptr<mlir::Pass> fir::createAffineDemotionPass() { |
| 181 | return std::make_unique<AffineDialectDemotion>(); |
| 182 | } |
| 183 | |