1 //===- BufferizableOpInterfaceImpl.cpp - Impl. of BufferizableOpInterface -===//
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 #include "mlir/Dialect/Arithmetic/Transforms/BufferizableOpInterfaceImpl.h"
10 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
11 #include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h"
12 #include "mlir/Dialect/Bufferization/Transforms/BufferUtils.h"
13 #include "mlir/Dialect/MemRef/IR/MemRef.h"
14 #include "mlir/IR/Dialect.h"
15 #include "mlir/IR/Operation.h"
16 
17 using namespace mlir;
18 using namespace mlir::bufferization;
19 
20 namespace {
21 /// Bufferization of arith.constant. Replace with memref.get_global.
22 struct ConstantOpInterface
23     : public BufferizableOpInterface::ExternalModel<ConstantOpInterface,
24                                                     arith::ConstantOp> {
25   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
26                           BufferizationState &state) const {
27     auto constantOp = cast<arith::ConstantOp>(op);
28 
29     // Only ranked tensors are supported.
30     if (!constantOp.getType().isa<RankedTensorType>())
31       return failure();
32 
33     // Only constants inside a module are supported.
34     auto moduleOp = constantOp->getParentOfType<ModuleOp>();
35     if (!moduleOp)
36       return failure();
37 
38     // Create global memory segment and replace tensor with memref pointing to
39     // that memory segment.
40     FailureOr<memref::GlobalOp> globalOp =
41         getGlobalFor(constantOp, state.getOptions().bufferAlignment);
42     if (failed(globalOp))
43       return failure();
44     memref::GlobalOp globalMemref = globalOp.getValue();
45     replaceOpWithNewBufferizedOp<memref::GetGlobalOp>(
46         rewriter, op, globalMemref.type(), globalMemref.getName());
47 
48     return success();
49   }
50 
51   bool isWritable(Operation *op, Value value,
52                   const AnalysisState &state) const {
53     // Memory locations returned by memref::GetGlobalOp may not be written to.
54     assert(value.isa<OpResult>());
55     return false;
56   }
57 };
58 
59 struct IndexCastOpInterface
60     : public BufferizableOpInterface::ExternalModel<IndexCastOpInterface,
61                                                     arith::IndexCastOp> {
62   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
63                               const AnalysisState &state) const {
64     return false;
65   }
66 
67   bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
68                                const AnalysisState &state) const {
69     return false;
70   }
71 
72   SmallVector<OpResult> getAliasingOpResult(Operation *op, OpOperand &opOperand,
73                                             const AnalysisState &state) const {
74     return {op->getResult(0)};
75   }
76 
77   BufferRelation bufferRelation(Operation *op, OpResult opResult,
78                                 const AnalysisState &state) const {
79     return BufferRelation::Equivalent;
80   }
81 
82   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
83                           BufferizationState &state) const {
84     auto castOp = cast<arith::IndexCastOp>(op);
85     auto resultTensorType = castOp.getType().cast<TensorType>();
86 
87     Value source = state.getBuffer(rewriter, castOp.getIn());
88     auto sourceType = source.getType().cast<BaseMemRefType>();
89 
90     // Result type should have same layout and address space as the source type.
91     BaseMemRefType resultType;
92     if (auto rankedMemRefType = sourceType.dyn_cast<MemRefType>()) {
93       resultType = MemRefType::get(
94           rankedMemRefType.getShape(), resultTensorType.getElementType(),
95           rankedMemRefType.getLayout(), rankedMemRefType.getMemorySpace());
96     } else {
97       auto unrankedMemrefType = sourceType.cast<UnrankedMemRefType>();
98       resultType = UnrankedMemRefType::get(resultTensorType.getElementType(),
99                                            unrankedMemrefType.getMemorySpace());
100     }
101 
102     replaceOpWithNewBufferizedOp<arith::IndexCastOp>(rewriter, op, resultType,
103                                                      source);
104     return success();
105   }
106 };
107 
108 /// Bufferization of arith.select. Just replace the operands.
109 struct SelectOpInterface
110     : public BufferizableOpInterface::ExternalModel<SelectOpInterface,
111                                                     arith::SelectOp> {
112   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
113                               const AnalysisState &state) const {
114     return false;
115   }
116 
117   bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
118                                const AnalysisState &state) const {
119     return false;
120   }
121 
122   SmallVector<OpResult> getAliasingOpResult(Operation *op, OpOperand &opOperand,
123                                             const AnalysisState &state) const {
124     return {op->getOpResult(0) /*result*/};
125   }
126 
127   SmallVector<OpOperand *>
128   getAliasingOpOperand(Operation *op, OpResult opResult,
129                        const AnalysisState &state) const {
130     return {&op->getOpOperand(1) /*true_value*/,
131             &op->getOpOperand(2) /*false_value*/};
132   }
133 
134   LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
135                           BufferizationState &state) const {
136     auto selectOp = cast<arith::SelectOp>(op);
137     Location loc = selectOp.getLoc();
138 
139     // TODO: It would be more efficient to copy the result of the `select` op
140     // instead of its OpOperands. In the worst case, 2 copies are inserted at
141     // the moment (one for each tensor). When copying the op result, only one
142     // copy would be needed.
143     Value trueBuffer = state.getBuffer(rewriter, selectOp.getTrueValue());
144     Value falseBuffer = state.getBuffer(rewriter, selectOp.getFalseValue());
145 
146     // The "true" and the "false" operands must have the same type. If the
147     // buffers have different types, they differ only in their layout map. Cast
148     // both of them to the most dynamic MemRef type.
149     if (trueBuffer.getType() != falseBuffer.getType()) {
150       auto trueType = trueBuffer.getType().cast<MemRefType>();
151       int64_t dynamicOffset = ShapedType::kDynamicStrideOrOffset;
152       SmallVector<int64_t> dynamicStrides(trueType.getRank(),
153                                           ShapedType::kDynamicStrideOrOffset);
154       AffineMap stridedLayout = makeStridedLinearLayoutMap(
155           dynamicStrides, dynamicOffset, op->getContext());
156       auto castedType =
157           MemRefType::get(trueType.getShape(), trueType.getElementType(),
158                           stridedLayout, trueType.getMemorySpaceAsInt());
159       trueBuffer = rewriter.create<memref::CastOp>(loc, castedType, trueBuffer);
160       falseBuffer =
161           rewriter.create<memref::CastOp>(loc, castedType, falseBuffer);
162     }
163 
164     replaceOpWithNewBufferizedOp<arith::SelectOp>(
165         rewriter, op, selectOp.getCondition(), trueBuffer, falseBuffer);
166     return success();
167   }
168 
169   BufferRelation bufferRelation(Operation *op, OpResult opResult,
170                                 const AnalysisState &state) const {
171     return BufferRelation::None;
172   }
173 };
174 
175 } // namespace
176 
177 void mlir::arith::registerBufferizableOpInterfaceExternalModels(
178     DialectRegistry &registry) {
179   registry.addExtension(+[](MLIRContext *ctx, ArithmeticDialect *dialect) {
180     ConstantOp::attachInterface<ConstantOpInterface>(*ctx);
181     IndexCastOp::attachInterface<IndexCastOpInterface>(*ctx);
182     SelectOp::attachInterface<SelectOpInterface>(*ctx);
183   });
184 }
185