1 //===- Bufferize.cpp - Bufferization utilities ----------------------------===// 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 "PassDetail.h" 10 11 #include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h" 12 #include "mlir/Dialect/Bufferization/IR/Bufferization.h" 13 #include "mlir/Dialect/Bufferization/Transforms/Bufferize.h" 14 #include "mlir/Dialect/Bufferization/Transforms/OneShotAnalysis.h" 15 #include "mlir/Dialect/Bufferization/Transforms/OneShotModuleBufferize.h" 16 #include "mlir/Dialect/Bufferization/Transforms/Passes.h" 17 #include "mlir/Dialect/Bufferization/Transforms/TensorCopyInsertion.h" 18 #include "mlir/Dialect/Func/IR/FuncOps.h" 19 #include "mlir/Dialect/MemRef/IR/MemRef.h" 20 #include "mlir/IR/Operation.h" 21 #include "mlir/Pass/PassManager.h" 22 #include "mlir/Transforms/GreedyPatternRewriteDriver.h" 23 #include "mlir/Transforms/Passes.h" 24 25 using namespace mlir; 26 using namespace mlir::bufferization; 27 28 //===----------------------------------------------------------------------===// 29 // BufferizeTypeConverter 30 //===----------------------------------------------------------------------===// 31 32 static Value materializeToTensor(OpBuilder &builder, TensorType type, 33 ValueRange inputs, Location loc) { 34 assert(inputs.size() == 1); 35 assert(inputs[0].getType().isa<BaseMemRefType>()); 36 return builder.create<bufferization::ToTensorOp>(loc, type, inputs[0]); 37 } 38 39 /// Registers conversions into BufferizeTypeConverter 40 BufferizeTypeConverter::BufferizeTypeConverter() { 41 // Keep all types unchanged. 42 addConversion([](Type type) { return type; }); 43 // Convert RankedTensorType to MemRefType. 44 addConversion([](RankedTensorType type) -> Type { 45 return MemRefType::get(type.getShape(), type.getElementType()); 46 }); 47 // Convert UnrankedTensorType to UnrankedMemRefType. 48 addConversion([](UnrankedTensorType type) -> Type { 49 return UnrankedMemRefType::get(type.getElementType(), 0); 50 }); 51 addArgumentMaterialization(materializeToTensor); 52 addSourceMaterialization(materializeToTensor); 53 addTargetMaterialization([](OpBuilder &builder, BaseMemRefType type, 54 ValueRange inputs, Location loc) -> Value { 55 assert(inputs.size() == 1 && "expected exactly one input"); 56 57 if (auto inputType = inputs[0].getType().dyn_cast<MemRefType>()) { 58 // MemRef to MemRef cast. 59 assert(inputType != type && "expected different types"); 60 // Unranked to ranked and ranked to unranked casts must be explicit. 61 auto rankedDestType = type.dyn_cast<MemRefType>(); 62 if (!rankedDestType) 63 return nullptr; 64 FailureOr<Value> replacement = 65 castOrReallocMemRefValue(builder, inputs[0], rankedDestType); 66 if (failed(replacement)) 67 return nullptr; 68 return *replacement; 69 } 70 71 if (inputs[0].getType().isa<TensorType>()) { 72 // Tensor to MemRef cast. 73 return builder.create<bufferization::ToMemrefOp>(loc, type, inputs[0]); 74 } 75 76 llvm_unreachable("only tensor/memref input types supported"); 77 }); 78 } 79 80 void mlir::bufferization::populateBufferizeMaterializationLegality( 81 ConversionTarget &target) { 82 target.addLegalOp<bufferization::ToTensorOp, bufferization::ToMemrefOp>(); 83 } 84 85 namespace { 86 // In a finalizing bufferize conversion, we know that all tensors have been 87 // converted to memrefs, thus, this op becomes an identity. 88 class BufferizeToTensorOp 89 : public OpConversionPattern<bufferization::ToTensorOp> { 90 public: 91 using OpConversionPattern::OpConversionPattern; 92 LogicalResult 93 matchAndRewrite(bufferization::ToTensorOp op, OpAdaptor adaptor, 94 ConversionPatternRewriter &rewriter) const override { 95 rewriter.replaceOp(op, adaptor.getMemref()); 96 return success(); 97 } 98 }; 99 } // namespace 100 101 namespace { 102 // In a finalizing bufferize conversion, we know that all tensors have been 103 // converted to memrefs, thus, this op becomes an identity. 104 class BufferizeToMemrefOp 105 : public OpConversionPattern<bufferization::ToMemrefOp> { 106 public: 107 using OpConversionPattern::OpConversionPattern; 108 LogicalResult 109 matchAndRewrite(bufferization::ToMemrefOp op, OpAdaptor adaptor, 110 ConversionPatternRewriter &rewriter) const override { 111 rewriter.replaceOp(op, adaptor.getTensor()); 112 return success(); 113 } 114 }; 115 } // namespace 116 117 void mlir::bufferization::populateEliminateBufferizeMaterializationsPatterns( 118 BufferizeTypeConverter &typeConverter, RewritePatternSet &patterns) { 119 patterns.add<BufferizeToTensorOp, BufferizeToMemrefOp>(typeConverter, 120 patterns.getContext()); 121 } 122 123 namespace { 124 struct FinalizingBufferizePass 125 : public FinalizingBufferizeBase<FinalizingBufferizePass> { 126 using FinalizingBufferizeBase< 127 FinalizingBufferizePass>::FinalizingBufferizeBase; 128 129 void runOnOperation() override { 130 auto func = getOperation(); 131 auto *context = &getContext(); 132 133 BufferizeTypeConverter typeConverter; 134 RewritePatternSet patterns(context); 135 ConversionTarget target(*context); 136 137 populateEliminateBufferizeMaterializationsPatterns(typeConverter, patterns); 138 139 // If all result types are legal, and all block arguments are legal (ensured 140 // by func conversion above), then all types in the program are legal. 141 // 142 // We also check that the operand types are legal to avoid creating invalid 143 // IR. For example, this prevents 144 // populateEliminateBufferizeMaterializationsPatterns from updating the 145 // types of the operands to a return op without updating the enclosing 146 // function. 147 target.markUnknownOpDynamicallyLegal( 148 [&](Operation *op) { return typeConverter.isLegal(op); }); 149 150 if (failed(applyFullConversion(func, target, std::move(patterns)))) 151 signalPassFailure(); 152 } 153 }; 154 155 static BufferizationOptions::LayoutMapOption 156 parseLayoutMapOption(const std::string &s) { 157 if (s == "fully-dynamic-layout-map") 158 return BufferizationOptions::LayoutMapOption::FullyDynamicLayoutMap; 159 if (s == "identity-layout-map") 160 return BufferizationOptions::LayoutMapOption::IdentityLayoutMap; 161 if (s == "infer-layout-map") 162 return BufferizationOptions::LayoutMapOption::InferLayoutMap; 163 llvm_unreachable("invalid layout map option"); 164 } 165 166 struct OneShotBufferizePass 167 : public OneShotBufferizeBase<OneShotBufferizePass> { 168 OneShotBufferizePass() : OneShotBufferizeBase<OneShotBufferizePass>() {} 169 170 explicit OneShotBufferizePass(const OneShotBufferizationOptions &options) 171 : options(options) {} 172 173 void getDependentDialects(DialectRegistry ®istry) const override { 174 registry 175 .insert<bufferization::BufferizationDialect, memref::MemRefDialect>(); 176 registerAllocationOpInterfaceExternalModels(registry); 177 } 178 179 void runOnOperation() override { 180 OneShotBufferizationOptions opt; 181 if (!options) { 182 // Make new bufferization options if none were provided when creating the 183 // pass. 184 opt.allowReturnAllocs = allowReturnAllocs; 185 opt.allowUnknownOps = allowUnknownOps; 186 opt.analysisFuzzerSeed = analysisFuzzerSeed; 187 opt.createDeallocs = createDeallocs; 188 opt.functionBoundaryTypeConversion = 189 parseLayoutMapOption(functionBoundaryTypeConversion); 190 if (mustInferMemorySpace) 191 opt.defaultMemorySpace = None; 192 opt.printConflicts = printConflicts; 193 opt.testAnalysisOnly = testAnalysisOnly; 194 opt.bufferizeFunctionBoundaries = bufferizeFunctionBoundaries; 195 196 // Configure type converter. 197 BufferizationOptions::LayoutMapOption unknownTypeConversionOption = 198 parseLayoutMapOption(unknownTypeConversion); 199 opt.unknownTypeConverterFn = [=](Value value, unsigned memorySpace, 200 const BufferizationOptions &options) { 201 auto tensorType = value.getType().cast<TensorType>(); 202 if (unknownTypeConversionOption == 203 BufferizationOptions::LayoutMapOption::IdentityLayoutMap) 204 return bufferization::getMemRefTypeWithStaticIdentityLayout( 205 tensorType, memorySpace); 206 assert( 207 unknownTypeConversionOption == 208 BufferizationOptions::LayoutMapOption::FullyDynamicLayoutMap && 209 "invalid layout map option"); 210 return bufferization::getMemRefTypeWithFullyDynamicLayout(tensorType, 211 memorySpace); 212 }; 213 214 // Configure op filter. 215 OpFilter::Entry::FilterFn filterFn = [&](Operation *op) { 216 // Filter may be specified via options. 217 if (this->dialectFilter.hasValue()) 218 return llvm::is_contained(this->dialectFilter, 219 op->getDialect()->getNamespace()); 220 // No filter specified: All other ops are allowed. 221 return true; 222 }; 223 opt.opFilter.allowOperation(filterFn); 224 } else { 225 opt = *options; 226 } 227 228 ModuleOp moduleOp = getOperation(); 229 if (opt.bufferizeFunctionBoundaries) { 230 if (failed(runOneShotModuleBufferize(moduleOp, opt))) { 231 signalPassFailure(); 232 return; 233 } 234 } else { 235 if (failed(runOneShotBufferize(moduleOp, opt))) { 236 signalPassFailure(); 237 return; 238 } 239 } 240 241 if (opt.testAnalysisOnly) 242 return; 243 244 OpPassManager cleanupPipeline("builtin.module"); 245 cleanupPipeline.addPass(createCanonicalizerPass()); 246 cleanupPipeline.addPass(createCSEPass()); 247 cleanupPipeline.addPass(createLoopInvariantCodeMotionPass()); 248 (void)runPipeline(cleanupPipeline, moduleOp); 249 } 250 251 private: 252 llvm::Optional<OneShotBufferizationOptions> options; 253 }; 254 } // namespace 255 256 namespace { 257 struct BufferizationBufferizePass 258 : public BufferizationBufferizeBase<BufferizationBufferizePass> { 259 void runOnOperation() override { 260 BufferizationOptions options = getPartialBufferizationOptions(); 261 options.opFilter.allowDialect<BufferizationDialect>(); 262 263 if (failed(bufferizeOp(getOperation(), options))) 264 signalPassFailure(); 265 } 266 267 void getDependentDialects(DialectRegistry ®istry) const override { 268 registry 269 .insert<bufferization::BufferizationDialect, memref::MemRefDialect>(); 270 } 271 }; 272 } // namespace 273 274 std::unique_ptr<Pass> mlir::bufferization::createBufferizationBufferizePass() { 275 return std::make_unique<BufferizationBufferizePass>(); 276 } 277 278 std::unique_ptr<Pass> mlir::bufferization::createOneShotBufferizePass() { 279 return std::make_unique<OneShotBufferizePass>(); 280 } 281 282 std::unique_ptr<Pass> mlir::bufferization::createOneShotBufferizePass( 283 const OneShotBufferizationOptions &options) { 284 return std::make_unique<OneShotBufferizePass>(options); 285 } 286 287 std::unique_ptr<OperationPass<func::FuncOp>> 288 mlir::bufferization::createFinalizingBufferizePass() { 289 return std::make_unique<FinalizingBufferizePass>(); 290 } 291 292 //===----------------------------------------------------------------------===// 293 // BufferizableOpInterface-based Bufferization 294 //===----------------------------------------------------------------------===// 295 296 static bool isaTensor(Type t) { return t.isa<TensorType>(); } 297 298 /// Return true if the given op has a tensor result or a tensor operand. 299 static bool hasTensorSemantics(Operation *op) { 300 if (auto funcOp = dyn_cast<FunctionOpInterface>(op)) { 301 bool hasTensorArg = any_of(funcOp.getArgumentTypes(), isaTensor); 302 bool hasTensorResult = any_of(funcOp.getResultTypes(), isaTensor); 303 return hasTensorArg || hasTensorResult; 304 } 305 306 bool hasTensorResult = any_of(op->getResultTypes(), isaTensor); 307 bool hasTensorOperand = any_of(op->getOperandTypes(), isaTensor); 308 return hasTensorResult || hasTensorOperand; 309 } 310 311 namespace { 312 /// A rewriter that keeps track of extra information during bufferization. 313 class BufferizationRewriter : public IRRewriter { 314 public: 315 BufferizationRewriter(MLIRContext *ctx, DenseSet<Operation *> &erasedOps, 316 DenseSet<Operation *> &toMemrefOps, 317 SmallVector<Operation *> &worklist, 318 const BufferizationOptions &options, 319 const OpFilter *opFilter) 320 : IRRewriter(ctx), erasedOps(erasedOps), toMemrefOps(toMemrefOps), 321 worklist(worklist), analysisState(options), opFilter(opFilter) {} 322 323 protected: 324 void notifyOperationRemoved(Operation *op) override { 325 IRRewriter::notifyOperationRemoved(op); 326 erasedOps.insert(op); 327 // Erase if present. 328 toMemrefOps.erase(op); 329 } 330 331 void notifyOperationInserted(Operation *op) override { 332 IRRewriter::notifyOperationInserted(op); 333 erasedOps.erase(op); 334 335 // Keep track of to_memref ops. 336 if (isa<ToMemrefOp>(op)) { 337 toMemrefOps.insert(op); 338 return; 339 } 340 341 // Skip to_tensor ops. 342 if (isa<ToTensorOp>(op)) 343 return; 344 345 // Skip non-tensor ops. 346 if (!hasTensorSemantics(op)) 347 return; 348 349 // Skip ops that are not allowed to be bufferized. 350 auto const &options = analysisState.getOptions(); 351 if (!options.isOpAllowed(op) || (opFilter && !opFilter->isOpAllowed(op))) 352 return; 353 354 #ifndef NDEBUG 355 // Read-only tensor ops may be created during bufferization. Ops that are 356 // writing should not be created because such ops were never analyzed. 357 // Bufferizing such ops could introduce a RaW conflict. 358 for (OpOperand &operand : op->getOpOperands()) 359 if (operand.get().getType().isa<TensorType>()) 360 assert(!analysisState.bufferizesToMemoryWrite(operand) && 361 "creating tensor ops that bufferize to a memory write is not " 362 "allowed during bufferization"); 363 #endif // NDEBUG 364 365 // Add op to worklist. 366 worklist.push_back(op); 367 } 368 369 private: 370 /// A set of all erased ops. 371 DenseSet<Operation *> &erasedOps; 372 373 /// A set of all to_memref ops. 374 DenseSet<Operation *> &toMemrefOps; 375 376 /// The worklist of ops to be bufferized. 377 SmallVector<Operation *> &worklist; 378 379 /// The analysis state. Used for debug assertions and access to the 380 /// bufferization options. 381 const AnalysisState analysisState; 382 383 /// An extra op filter for bufferization. 384 const OpFilter *opFilter; 385 }; 386 } // namespace 387 388 LogicalResult bufferization::bufferizeOp(Operation *op, 389 const BufferizationOptions &options, 390 bool copyBeforeWrite, 391 const OpFilter *opFilter) { 392 if (copyBeforeWrite) { 393 AnalysisState state(options); 394 if (failed(insertTensorCopies(op, state))) 395 return failure(); 396 } 397 398 // Keep track of to_memref ops. 399 DenseSet<Operation *> toMemrefOps; 400 op->walk([&](ToMemrefOp toMemrefOp) { toMemrefOps.insert(toMemrefOp); }); 401 402 // Gather all bufferizable ops in top-to-bottom order. 403 // 404 // We should ideally know the exact memref type of all operands when 405 // bufferizing an op. (This is the case when bufferizing top-to-bottom.) 406 // Otherwise, we have to use a memref type with a fully dynamic layout map to 407 // avoid copies. We are currently missing patterns for layout maps to 408 // canonicalize away (or canonicalize to more precise layouts). 409 // 410 // FuncOps must be bufferized before their bodies, so add them to the worklist 411 // first. 412 SmallVector<Operation *> worklist; 413 op->walk([&](func::FuncOp funcOp) { 414 if (hasTensorSemantics(funcOp)) 415 worklist.push_back(funcOp); 416 }); 417 op->walk<WalkOrder::PostOrder>([&](Operation *op) { 418 if (hasTensorSemantics(op) && !isa<func::FuncOp>(op)) 419 worklist.push_back(op); 420 }); 421 422 // Keep track of all erased ops. 423 DenseSet<Operation *> erasedOps; 424 425 // Bufferize all ops. 426 BufferizationRewriter rewriter(op->getContext(), erasedOps, toMemrefOps, 427 worklist, options, opFilter); 428 for (unsigned i = 0; i < worklist.size(); ++i) { 429 Operation *op = worklist[i]; 430 // Skip ops that were erased. 431 if (erasedOps.contains(op)) 432 continue; 433 // Skip ops that are not bufferizable or not allowed. 434 auto bufferizableOp = options.dynCastBufferizableOp(op); 435 if (!bufferizableOp) 436 continue; 437 if (opFilter && !opFilter->isOpAllowed(op)) 438 continue; 439 // Skip ops that no longer have tensor semantics. 440 if (!hasTensorSemantics(op)) 441 continue; 442 // Bufferize the op. 443 rewriter.setInsertionPoint(op); 444 if (failed(bufferizableOp.bufferize(rewriter, options))) 445 return op->emitError("failed to bufferize op"); 446 } 447 448 // Fold all to_memref(to_tensor(x)) pairs. 449 for (Operation *op : toMemrefOps) { 450 rewriter.setInsertionPoint(op); 451 (void)bufferization::foldToMemrefToTensorPair(rewriter, 452 cast<ToMemrefOp>(op)); 453 } 454 455 /// Check the result of bufferization. Return an error if an op was not 456 /// bufferized, unless partial bufferization is allowed. 457 if (options.allowUnknownOps) 458 return success(); 459 460 for (Operation *op : worklist) { 461 // Skip ops that are entirely gone. 462 if (erasedOps.contains(op)) 463 continue; 464 // Ops that no longer have tensor semantics (because they were updated 465 // in-place) are allowed. 466 if (!hasTensorSemantics(op)) 467 continue; 468 // Continue ops that are not allowed. 469 if (!options.isOpAllowed(op)) 470 continue; 471 if (opFilter && !opFilter->isOpAllowed(op)) 472 continue; 473 // Ops without any uses and no side effects will fold away. 474 if (op->getUses().empty() && MemoryEffectOpInterface::hasNoEffect(op)) 475 continue; 476 // ToTensorOps/ToMemrefOps are allowed in the output. 477 if (isa<ToTensorOp, ToMemrefOp>(op)) 478 continue; 479 return op->emitError("op was not bufferized"); 480 } 481 482 return success(); 483 } 484 485 BufferizationOptions bufferization::getPartialBufferizationOptions() { 486 BufferizationOptions options; 487 options.allowUnknownOps = true; 488 options.createDeallocs = false; 489 options.enforceAliasingInvariants = false; 490 options.unknownTypeConverterFn = [](Value value, unsigned memorySpace, 491 const BufferizationOptions &options) { 492 return getMemRefTypeWithStaticIdentityLayout( 493 value.getType().cast<TensorType>(), memorySpace); 494 }; 495 options.opFilter.allowDialect<BufferizationDialect>(); 496 return options; 497 } 498