1 //===-- BoxedProcedure.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 #include "flang/Optimizer/CodeGen/CodeGen.h" 10 11 #include "flang/Optimizer/Builder/FIRBuilder.h" 12 #include "flang/Optimizer/Builder/LowLevelIntrinsics.h" 13 #include "flang/Optimizer/Dialect/FIRDialect.h" 14 #include "flang/Optimizer/Dialect/FIROps.h" 15 #include "flang/Optimizer/Dialect/FIRType.h" 16 #include "flang/Optimizer/Dialect/Support/FIRContext.h" 17 #include "flang/Optimizer/Support/FatalError.h" 18 #include "flang/Optimizer/Support/InternalNames.h" 19 #include "mlir/IR/PatternMatch.h" 20 #include "mlir/Pass/Pass.h" 21 #include "mlir/Transforms/DialectConversion.h" 22 #include "llvm/ADT/DenseMap.h" 23 24 namespace fir { 25 #define GEN_PASS_DEF_BOXEDPROCEDUREPASS 26 #include "flang/Optimizer/CodeGen/CGPasses.h.inc" 27 } // namespace fir 28 29 #define DEBUG_TYPE "flang-procedure-pointer" 30 31 using namespace fir; 32 33 namespace { 34 /// Options to the procedure pointer pass. 35 struct BoxedProcedureOptions { 36 // Lower the boxproc abstraction to function pointers and thunks where 37 // required. 38 bool useThunks = true; 39 }; 40 41 /// This type converter rewrites all `!fir.boxproc<Func>` types to `Func` types. 42 class BoxprocTypeRewriter : public mlir::TypeConverter { 43 public: 44 using mlir::TypeConverter::convertType; 45 46 /// Does the type \p ty need to be converted? 47 /// Any type that is a `!fir.boxproc` in whole or in part will need to be 48 /// converted to a function type to lower the IR to function pointer form in 49 /// the default implementation performed in this pass. Other implementations 50 /// are possible, so those may convert `!fir.boxproc` to some other type or 51 /// not at all depending on the implementation target's characteristics and 52 /// preference. 53 bool needsConversion(mlir::Type ty) { 54 if (ty.isa<BoxProcType>()) 55 return true; 56 if (auto funcTy = ty.dyn_cast<mlir::FunctionType>()) { 57 for (auto t : funcTy.getInputs()) 58 if (needsConversion(t)) 59 return true; 60 for (auto t : funcTy.getResults()) 61 if (needsConversion(t)) 62 return true; 63 return false; 64 } 65 if (auto tupleTy = ty.dyn_cast<mlir::TupleType>()) { 66 for (auto t : tupleTy.getTypes()) 67 if (needsConversion(t)) 68 return true; 69 return false; 70 } 71 if (auto recTy = ty.dyn_cast<RecordType>()) { 72 auto visited = visitedTypes.find(ty); 73 if (visited != visitedTypes.end()) 74 return visited->second; 75 [[maybe_unused]] auto newIt = visitedTypes.try_emplace(ty, false); 76 assert(newIt.second && "expected ty to not be in the map"); 77 bool wasAlreadyVisitingRecordType = needConversionIsVisitingRecordType; 78 needConversionIsVisitingRecordType = true; 79 bool result = false; 80 for (auto t : recTy.getTypeList()) { 81 if (needsConversion(t.second)) { 82 result = true; 83 break; 84 } 85 } 86 // Only keep the result cached if the fir.type visited was a "top-level 87 // type". Nested types with a recursive reference to the "top-level type" 88 // may incorrectly have been resolved as not needed conversions because it 89 // had not been determined yet if the "top-level type" needed conversion. 90 // This is not an issue to determine the "top-level type" need of 91 // conversion, but the result should not be kept and later used in other 92 // contexts. 93 needConversionIsVisitingRecordType = wasAlreadyVisitingRecordType; 94 if (needConversionIsVisitingRecordType) 95 visitedTypes.erase(ty); 96 else 97 visitedTypes.find(ty)->second = result; 98 return result; 99 } 100 if (auto boxTy = ty.dyn_cast<BaseBoxType>()) 101 return needsConversion(boxTy.getEleTy()); 102 if (isa_ref_type(ty)) 103 return needsConversion(unwrapRefType(ty)); 104 if (auto t = ty.dyn_cast<SequenceType>()) 105 return needsConversion(unwrapSequenceType(ty)); 106 return false; 107 } 108 109 BoxprocTypeRewriter(mlir::Location location) : loc{location} { 110 addConversion([](mlir::Type ty) { return ty; }); 111 addConversion( 112 [&](BoxProcType boxproc) { return convertType(boxproc.getEleTy()); }); 113 addConversion([&](mlir::TupleType tupTy) { 114 llvm::SmallVector<mlir::Type> memTys; 115 for (auto ty : tupTy.getTypes()) 116 memTys.push_back(convertType(ty)); 117 return mlir::TupleType::get(tupTy.getContext(), memTys); 118 }); 119 addConversion([&](mlir::FunctionType funcTy) { 120 llvm::SmallVector<mlir::Type> inTys; 121 llvm::SmallVector<mlir::Type> resTys; 122 for (auto ty : funcTy.getInputs()) 123 inTys.push_back(convertType(ty)); 124 for (auto ty : funcTy.getResults()) 125 resTys.push_back(convertType(ty)); 126 return mlir::FunctionType::get(funcTy.getContext(), inTys, resTys); 127 }); 128 addConversion([&](ReferenceType ty) { 129 return ReferenceType::get(convertType(ty.getEleTy())); 130 }); 131 addConversion([&](PointerType ty) { 132 return PointerType::get(convertType(ty.getEleTy())); 133 }); 134 addConversion( 135 [&](HeapType ty) { return HeapType::get(convertType(ty.getEleTy())); }); 136 addConversion([&](fir::LLVMPointerType ty) { 137 return fir::LLVMPointerType::get(convertType(ty.getEleTy())); 138 }); 139 addConversion( 140 [&](BoxType ty) { return BoxType::get(convertType(ty.getEleTy())); }); 141 addConversion([&](ClassType ty) { 142 return ClassType::get(convertType(ty.getEleTy())); 143 }); 144 addConversion([&](SequenceType ty) { 145 // TODO: add ty.getLayoutMap() as needed. 146 return SequenceType::get(ty.getShape(), convertType(ty.getEleTy())); 147 }); 148 addConversion([&](RecordType ty) -> mlir::Type { 149 if (!needsConversion(ty)) 150 return ty; 151 if (auto converted = convertedTypes.lookup(ty)) 152 return converted; 153 auto rec = RecordType::get(ty.getContext(), 154 ty.getName().str() + boxprocSuffix.str()); 155 if (rec.isFinalized()) 156 return rec; 157 [[maybe_unused]] auto it = convertedTypes.try_emplace(ty, rec); 158 assert(it.second && "expected ty to not be in the map"); 159 std::vector<RecordType::TypePair> ps = ty.getLenParamList(); 160 std::vector<RecordType::TypePair> cs; 161 for (auto t : ty.getTypeList()) { 162 if (needsConversion(t.second)) 163 cs.emplace_back(t.first, convertType(t.second)); 164 else 165 cs.emplace_back(t.first, t.second); 166 } 167 rec.finalize(ps, cs); 168 return rec; 169 }); 170 addArgumentMaterialization(materializeProcedure); 171 addSourceMaterialization(materializeProcedure); 172 addTargetMaterialization(materializeProcedure); 173 } 174 175 static mlir::Value materializeProcedure(mlir::OpBuilder &builder, 176 BoxProcType type, 177 mlir::ValueRange inputs, 178 mlir::Location loc) { 179 assert(inputs.size() == 1); 180 return builder.create<ConvertOp>(loc, unwrapRefType(type.getEleTy()), 181 inputs[0]); 182 } 183 184 void setLocation(mlir::Location location) { loc = location; } 185 186 private: 187 // Maps to deal with recursive derived types (avoid infinite loops). 188 // Caching is also beneficial for apps with big types (dozens of 189 // components and or parent types), so the lifetime of the cache 190 // is the whole pass. 191 llvm::DenseMap<mlir::Type, bool> visitedTypes; 192 bool needConversionIsVisitingRecordType = false; 193 llvm::DenseMap<mlir::Type, mlir::Type> convertedTypes; 194 mlir::Location loc; 195 }; 196 197 /// A `boxproc` is an abstraction for a Fortran procedure reference. Typically, 198 /// Fortran procedures can be referenced directly through a function pointer. 199 /// However, Fortran has one-level dynamic scoping between a host procedure and 200 /// its internal procedures. This allows internal procedures to directly access 201 /// and modify the state of the host procedure's variables. 202 /// 203 /// There are any number of possible implementations possible. 204 /// 205 /// The implementation used here is to convert `boxproc` values to function 206 /// pointers everywhere. If a `boxproc` value includes a frame pointer to the 207 /// host procedure's data, then a thunk will be created at runtime to capture 208 /// the frame pointer during execution. In LLVM IR, the frame pointer is 209 /// designated with the `nest` attribute. The thunk's address will then be used 210 /// as the call target instead of the original function's address directly. 211 class BoxedProcedurePass 212 : public fir::impl::BoxedProcedurePassBase<BoxedProcedurePass> { 213 public: 214 BoxedProcedurePass() { options = {true}; } 215 BoxedProcedurePass(bool useThunks) { options = {useThunks}; } 216 217 inline mlir::ModuleOp getModule() { return getOperation(); } 218 219 void runOnOperation() override final { 220 if (options.useThunks) { 221 auto *context = &getContext(); 222 mlir::IRRewriter rewriter(context); 223 BoxprocTypeRewriter typeConverter(mlir::UnknownLoc::get(context)); 224 mlir::Dialect *firDialect = context->getLoadedDialect("fir"); 225 getModule().walk([&](mlir::Operation *op) { 226 bool opIsValid = true; 227 typeConverter.setLocation(op->getLoc()); 228 if (auto addr = mlir::dyn_cast<BoxAddrOp>(op)) { 229 mlir::Type ty = addr.getVal().getType(); 230 mlir::Type resTy = addr.getResult().getType(); 231 if (llvm::isa<mlir::FunctionType>(ty) || 232 llvm::isa<fir::BoxProcType>(ty)) { 233 // Rewrite all `fir.box_addr` ops on values of type `!fir.boxproc` 234 // or function type to be `fir.convert` ops. 235 rewriter.setInsertionPoint(addr); 236 rewriter.replaceOpWithNewOp<ConvertOp>( 237 addr, typeConverter.convertType(addr.getType()), addr.getVal()); 238 opIsValid = false; 239 } else if (typeConverter.needsConversion(resTy)) { 240 rewriter.startOpModification(op); 241 op->getResult(0).setType(typeConverter.convertType(resTy)); 242 rewriter.finalizeOpModification(op); 243 } 244 } else if (auto func = mlir::dyn_cast<mlir::func::FuncOp>(op)) { 245 mlir::FunctionType ty = func.getFunctionType(); 246 if (typeConverter.needsConversion(ty)) { 247 rewriter.startOpModification(func); 248 auto toTy = 249 typeConverter.convertType(ty).cast<mlir::FunctionType>(); 250 if (!func.empty()) 251 for (auto e : llvm::enumerate(toTy.getInputs())) { 252 unsigned i = e.index(); 253 auto &block = func.front(); 254 block.insertArgument(i, e.value(), func.getLoc()); 255 block.getArgument(i + 1).replaceAllUsesWith( 256 block.getArgument(i)); 257 block.eraseArgument(i + 1); 258 } 259 func.setType(toTy); 260 rewriter.finalizeOpModification(func); 261 } 262 } else if (auto embox = mlir::dyn_cast<EmboxProcOp>(op)) { 263 // Rewrite all `fir.emboxproc` ops to either `fir.convert` or a thunk 264 // as required. 265 mlir::Type toTy = typeConverter.convertType( 266 embox.getType().cast<BoxProcType>().getEleTy()); 267 rewriter.setInsertionPoint(embox); 268 if (embox.getHost()) { 269 // Create the thunk. 270 auto module = embox->getParentOfType<mlir::ModuleOp>(); 271 FirOpBuilder builder(rewriter, module); 272 auto loc = embox.getLoc(); 273 mlir::Type i8Ty = builder.getI8Type(); 274 mlir::Type i8Ptr = builder.getRefType(i8Ty); 275 mlir::Type buffTy = SequenceType::get({32}, i8Ty); 276 auto buffer = builder.create<AllocaOp>(loc, buffTy); 277 mlir::Value closure = 278 builder.createConvert(loc, i8Ptr, embox.getHost()); 279 mlir::Value tramp = builder.createConvert(loc, i8Ptr, buffer); 280 mlir::Value func = 281 builder.createConvert(loc, i8Ptr, embox.getFunc()); 282 builder.create<fir::CallOp>( 283 loc, factory::getLlvmInitTrampoline(builder), 284 llvm::ArrayRef<mlir::Value>{tramp, func, closure}); 285 auto adjustCall = builder.create<fir::CallOp>( 286 loc, factory::getLlvmAdjustTrampoline(builder), 287 llvm::ArrayRef<mlir::Value>{tramp}); 288 rewriter.replaceOpWithNewOp<ConvertOp>(embox, toTy, 289 adjustCall.getResult(0)); 290 opIsValid = false; 291 } else { 292 // Just forward the function as a pointer. 293 rewriter.replaceOpWithNewOp<ConvertOp>(embox, toTy, 294 embox.getFunc()); 295 opIsValid = false; 296 } 297 } else if (auto global = mlir::dyn_cast<GlobalOp>(op)) { 298 auto ty = global.getType(); 299 if (typeConverter.needsConversion(ty)) { 300 rewriter.startOpModification(global); 301 auto toTy = typeConverter.convertType(ty); 302 global.setType(toTy); 303 rewriter.finalizeOpModification(global); 304 } 305 } else if (auto mem = mlir::dyn_cast<AllocaOp>(op)) { 306 auto ty = mem.getType(); 307 if (typeConverter.needsConversion(ty)) { 308 rewriter.setInsertionPoint(mem); 309 auto toTy = typeConverter.convertType(unwrapRefType(ty)); 310 bool isPinned = mem.getPinned(); 311 llvm::StringRef uniqName = 312 mem.getUniqName().value_or(llvm::StringRef()); 313 llvm::StringRef bindcName = 314 mem.getBindcName().value_or(llvm::StringRef()); 315 rewriter.replaceOpWithNewOp<AllocaOp>( 316 mem, toTy, uniqName, bindcName, isPinned, mem.getTypeparams(), 317 mem.getShape()); 318 opIsValid = false; 319 } 320 } else if (auto mem = mlir::dyn_cast<AllocMemOp>(op)) { 321 auto ty = mem.getType(); 322 if (typeConverter.needsConversion(ty)) { 323 rewriter.setInsertionPoint(mem); 324 auto toTy = typeConverter.convertType(unwrapRefType(ty)); 325 llvm::StringRef uniqName = 326 mem.getUniqName().value_or(llvm::StringRef()); 327 llvm::StringRef bindcName = 328 mem.getBindcName().value_or(llvm::StringRef()); 329 rewriter.replaceOpWithNewOp<AllocMemOp>( 330 mem, toTy, uniqName, bindcName, mem.getTypeparams(), 331 mem.getShape()); 332 opIsValid = false; 333 } 334 } else if (auto coor = mlir::dyn_cast<CoordinateOp>(op)) { 335 auto ty = coor.getType(); 336 mlir::Type baseTy = coor.getBaseType(); 337 if (typeConverter.needsConversion(ty) || 338 typeConverter.needsConversion(baseTy)) { 339 rewriter.setInsertionPoint(coor); 340 auto toTy = typeConverter.convertType(ty); 341 auto toBaseTy = typeConverter.convertType(baseTy); 342 rewriter.replaceOpWithNewOp<CoordinateOp>(coor, toTy, coor.getRef(), 343 coor.getCoor(), toBaseTy); 344 opIsValid = false; 345 } 346 } else if (auto index = mlir::dyn_cast<FieldIndexOp>(op)) { 347 auto ty = index.getType(); 348 mlir::Type onTy = index.getOnType(); 349 if (typeConverter.needsConversion(ty) || 350 typeConverter.needsConversion(onTy)) { 351 rewriter.setInsertionPoint(index); 352 auto toTy = typeConverter.convertType(ty); 353 auto toOnTy = typeConverter.convertType(onTy); 354 rewriter.replaceOpWithNewOp<FieldIndexOp>( 355 index, toTy, index.getFieldId(), toOnTy, index.getTypeparams()); 356 opIsValid = false; 357 } 358 } else if (auto index = mlir::dyn_cast<LenParamIndexOp>(op)) { 359 auto ty = index.getType(); 360 mlir::Type onTy = index.getOnType(); 361 if (typeConverter.needsConversion(ty) || 362 typeConverter.needsConversion(onTy)) { 363 rewriter.setInsertionPoint(index); 364 auto toTy = typeConverter.convertType(ty); 365 auto toOnTy = typeConverter.convertType(onTy); 366 rewriter.replaceOpWithNewOp<LenParamIndexOp>( 367 index, toTy, index.getFieldId(), toOnTy, index.getTypeparams()); 368 opIsValid = false; 369 } 370 } else if (op->getDialect() == firDialect) { 371 rewriter.startOpModification(op); 372 for (auto i : llvm::enumerate(op->getResultTypes())) 373 if (typeConverter.needsConversion(i.value())) { 374 auto toTy = typeConverter.convertType(i.value()); 375 op->getResult(i.index()).setType(toTy); 376 } 377 rewriter.finalizeOpModification(op); 378 } 379 // Ensure block arguments are updated if needed. 380 if (opIsValid && op->getNumRegions() != 0) { 381 rewriter.startOpModification(op); 382 for (mlir::Region ®ion : op->getRegions()) 383 for (mlir::Block &block : region.getBlocks()) 384 for (mlir::BlockArgument blockArg : block.getArguments()) 385 if (typeConverter.needsConversion(blockArg.getType())) { 386 mlir::Type toTy = 387 typeConverter.convertType(blockArg.getType()); 388 blockArg.setType(toTy); 389 } 390 rewriter.finalizeOpModification(op); 391 } 392 }); 393 } 394 } 395 396 private: 397 BoxedProcedureOptions options; 398 }; 399 } // namespace 400 401 std::unique_ptr<mlir::Pass> fir::createBoxedProcedurePass() { 402 return std::make_unique<BoxedProcedurePass>(); 403 } 404 405 std::unique_ptr<mlir::Pass> fir::createBoxedProcedurePass(bool useThunks) { 406 return std::make_unique<BoxedProcedurePass>(useThunks); 407 } 408