1 //===- ModuleTranslation.cpp - MLIR to LLVM conversion --------------------===// 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 file implements the translation between an MLIR LLVM dialect module and 10 // the corresponding LLVMIR module. It only handles core LLVM IR operations. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "mlir/Target/LLVMIR/ModuleTranslation.h" 15 16 #include "mlir/Dialect/LLVMIR/LLVMDialect.h" 17 #include "mlir/IR/Attributes.h" 18 #include "mlir/IR/Module.h" 19 #include "mlir/IR/StandardTypes.h" 20 #include "mlir/Support/LLVM.h" 21 22 #include "llvm/ADT/SetVector.h" 23 #include "llvm/IR/BasicBlock.h" 24 #include "llvm/IR/Constants.h" 25 #include "llvm/IR/DerivedTypes.h" 26 #include "llvm/IR/IRBuilder.h" 27 #include "llvm/IR/LLVMContext.h" 28 #include "llvm/IR/Module.h" 29 #include "llvm/Transforms/Utils/Cloning.h" 30 31 using namespace mlir; 32 using namespace mlir::LLVM; 33 34 #include "mlir/Dialect/LLVMIR/LLVMConversionEnumsToLLVM.inc" 35 36 /// Builds a constant of a sequential LLVM type `type`, potentially containing 37 /// other sequential types recursively, from the individual constant values 38 /// provided in `constants`. `shape` contains the number of elements in nested 39 /// sequential types. Reports errors at `loc` and returns nullptr on error. 40 static llvm::Constant * 41 buildSequentialConstant(ArrayRef<llvm::Constant *> &constants, 42 ArrayRef<int64_t> shape, llvm::Type *type, 43 Location loc) { 44 if (shape.empty()) { 45 llvm::Constant *result = constants.front(); 46 constants = constants.drop_front(); 47 return result; 48 } 49 50 if (!isa<llvm::SequentialType>(type)) { 51 emitError(loc) << "expected sequential LLVM types wrapping a scalar"; 52 return nullptr; 53 } 54 55 llvm::Type *elementType = type->getSequentialElementType(); 56 SmallVector<llvm::Constant *, 8> nested; 57 nested.reserve(shape.front()); 58 for (int64_t i = 0; i < shape.front(); ++i) { 59 nested.push_back(buildSequentialConstant(constants, shape.drop_front(), 60 elementType, loc)); 61 if (!nested.back()) 62 return nullptr; 63 } 64 65 if (shape.size() == 1 && type->isVectorTy()) 66 return llvm::ConstantVector::get(nested); 67 return llvm::ConstantArray::get( 68 llvm::ArrayType::get(elementType, shape.front()), nested); 69 } 70 71 /// Returns the first non-sequential type nested in sequential types. 72 static llvm::Type *getInnermostElementType(llvm::Type *type) { 73 while (isa<llvm::SequentialType>(type)) 74 type = type->getSequentialElementType(); 75 return type; 76 } 77 78 /// Create an LLVM IR constant of `llvmType` from the MLIR attribute `attr`. 79 /// This currently supports integer, floating point, splat and dense element 80 /// attributes and combinations thereof. In case of error, report it to `loc` 81 /// and return nullptr. 82 llvm::Constant *ModuleTranslation::getLLVMConstant(llvm::Type *llvmType, 83 Attribute attr, 84 Location loc) { 85 if (!attr) 86 return llvm::UndefValue::get(llvmType); 87 if (llvmType->isStructTy()) { 88 emitError(loc, "struct types are not supported in constants"); 89 return nullptr; 90 } 91 if (auto intAttr = attr.dyn_cast<IntegerAttr>()) 92 return llvm::ConstantInt::get(llvmType, intAttr.getValue()); 93 if (auto floatAttr = attr.dyn_cast<FloatAttr>()) 94 return llvm::ConstantFP::get(llvmType, floatAttr.getValue()); 95 if (auto funcAttr = attr.dyn_cast<FlatSymbolRefAttr>()) 96 return functionMapping.lookup(funcAttr.getValue()); 97 if (auto splatAttr = attr.dyn_cast<SplatElementsAttr>()) { 98 auto *sequentialType = cast<llvm::SequentialType>(llvmType); 99 auto elementType = sequentialType->getElementType(); 100 uint64_t numElements = sequentialType->getNumElements(); 101 // Splat value is a scalar. Extract it only if the element type is not 102 // another sequence type. The recursion terminates because each step removes 103 // one outer sequential type. 104 llvm::Constant *child = getLLVMConstant( 105 elementType, 106 isa<llvm::SequentialType>(elementType) ? splatAttr 107 : splatAttr.getSplatValue(), 108 loc); 109 if (!child) 110 return nullptr; 111 if (llvmType->isVectorTy()) 112 return llvm::ConstantVector::getSplat(numElements, child); 113 if (llvmType->isArrayTy()) { 114 auto arrayType = llvm::ArrayType::get(elementType, numElements); 115 SmallVector<llvm::Constant *, 8> constants(numElements, child); 116 return llvm::ConstantArray::get(arrayType, constants); 117 } 118 } 119 120 if (auto elementsAttr = attr.dyn_cast<ElementsAttr>()) { 121 assert(elementsAttr.getType().hasStaticShape()); 122 assert(elementsAttr.getNumElements() != 0 && 123 "unexpected empty elements attribute"); 124 assert(!elementsAttr.getType().getShape().empty() && 125 "unexpected empty elements attribute shape"); 126 127 SmallVector<llvm::Constant *, 8> constants; 128 constants.reserve(elementsAttr.getNumElements()); 129 llvm::Type *innermostType = getInnermostElementType(llvmType); 130 for (auto n : elementsAttr.getValues<Attribute>()) { 131 constants.push_back(getLLVMConstant(innermostType, n, loc)); 132 if (!constants.back()) 133 return nullptr; 134 } 135 ArrayRef<llvm::Constant *> constantsRef = constants; 136 llvm::Constant *result = buildSequentialConstant( 137 constantsRef, elementsAttr.getType().getShape(), llvmType, loc); 138 assert(constantsRef.empty() && "did not consume all elemental constants"); 139 return result; 140 } 141 142 if (auto stringAttr = attr.dyn_cast<StringAttr>()) { 143 return llvm::ConstantDataArray::get( 144 llvmModule->getContext(), ArrayRef<char>{stringAttr.getValue().data(), 145 stringAttr.getValue().size()}); 146 } 147 emitError(loc, "unsupported constant value"); 148 return nullptr; 149 } 150 151 /// Convert MLIR integer comparison predicate to LLVM IR comparison predicate. 152 static llvm::CmpInst::Predicate getLLVMCmpPredicate(ICmpPredicate p) { 153 switch (p) { 154 case LLVM::ICmpPredicate::eq: 155 return llvm::CmpInst::Predicate::ICMP_EQ; 156 case LLVM::ICmpPredicate::ne: 157 return llvm::CmpInst::Predicate::ICMP_NE; 158 case LLVM::ICmpPredicate::slt: 159 return llvm::CmpInst::Predicate::ICMP_SLT; 160 case LLVM::ICmpPredicate::sle: 161 return llvm::CmpInst::Predicate::ICMP_SLE; 162 case LLVM::ICmpPredicate::sgt: 163 return llvm::CmpInst::Predicate::ICMP_SGT; 164 case LLVM::ICmpPredicate::sge: 165 return llvm::CmpInst::Predicate::ICMP_SGE; 166 case LLVM::ICmpPredicate::ult: 167 return llvm::CmpInst::Predicate::ICMP_ULT; 168 case LLVM::ICmpPredicate::ule: 169 return llvm::CmpInst::Predicate::ICMP_ULE; 170 case LLVM::ICmpPredicate::ugt: 171 return llvm::CmpInst::Predicate::ICMP_UGT; 172 case LLVM::ICmpPredicate::uge: 173 return llvm::CmpInst::Predicate::ICMP_UGE; 174 } 175 llvm_unreachable("incorrect comparison predicate"); 176 } 177 178 static llvm::CmpInst::Predicate getLLVMCmpPredicate(FCmpPredicate p) { 179 switch (p) { 180 case LLVM::FCmpPredicate::_false: 181 return llvm::CmpInst::Predicate::FCMP_FALSE; 182 case LLVM::FCmpPredicate::oeq: 183 return llvm::CmpInst::Predicate::FCMP_OEQ; 184 case LLVM::FCmpPredicate::ogt: 185 return llvm::CmpInst::Predicate::FCMP_OGT; 186 case LLVM::FCmpPredicate::oge: 187 return llvm::CmpInst::Predicate::FCMP_OGE; 188 case LLVM::FCmpPredicate::olt: 189 return llvm::CmpInst::Predicate::FCMP_OLT; 190 case LLVM::FCmpPredicate::ole: 191 return llvm::CmpInst::Predicate::FCMP_OLE; 192 case LLVM::FCmpPredicate::one: 193 return llvm::CmpInst::Predicate::FCMP_ONE; 194 case LLVM::FCmpPredicate::ord: 195 return llvm::CmpInst::Predicate::FCMP_ORD; 196 case LLVM::FCmpPredicate::ueq: 197 return llvm::CmpInst::Predicate::FCMP_UEQ; 198 case LLVM::FCmpPredicate::ugt: 199 return llvm::CmpInst::Predicate::FCMP_UGT; 200 case LLVM::FCmpPredicate::uge: 201 return llvm::CmpInst::Predicate::FCMP_UGE; 202 case LLVM::FCmpPredicate::ult: 203 return llvm::CmpInst::Predicate::FCMP_ULT; 204 case LLVM::FCmpPredicate::ule: 205 return llvm::CmpInst::Predicate::FCMP_ULE; 206 case LLVM::FCmpPredicate::une: 207 return llvm::CmpInst::Predicate::FCMP_UNE; 208 case LLVM::FCmpPredicate::uno: 209 return llvm::CmpInst::Predicate::FCMP_UNO; 210 case LLVM::FCmpPredicate::_true: 211 return llvm::CmpInst::Predicate::FCMP_TRUE; 212 } 213 llvm_unreachable("incorrect comparison predicate"); 214 } 215 216 static llvm::AtomicRMWInst::BinOp getLLVMAtomicBinOp(AtomicBinOp op) { 217 switch (op) { 218 case LLVM::AtomicBinOp::xchg: 219 return llvm::AtomicRMWInst::BinOp::Xchg; 220 case LLVM::AtomicBinOp::add: 221 return llvm::AtomicRMWInst::BinOp::Add; 222 case LLVM::AtomicBinOp::sub: 223 return llvm::AtomicRMWInst::BinOp::Sub; 224 case LLVM::AtomicBinOp::_and: 225 return llvm::AtomicRMWInst::BinOp::And; 226 case LLVM::AtomicBinOp::nand: 227 return llvm::AtomicRMWInst::BinOp::Nand; 228 case LLVM::AtomicBinOp::_or: 229 return llvm::AtomicRMWInst::BinOp::Or; 230 case LLVM::AtomicBinOp::_xor: 231 return llvm::AtomicRMWInst::BinOp::Xor; 232 case LLVM::AtomicBinOp::max: 233 return llvm::AtomicRMWInst::BinOp::Max; 234 case LLVM::AtomicBinOp::min: 235 return llvm::AtomicRMWInst::BinOp::Min; 236 case LLVM::AtomicBinOp::umax: 237 return llvm::AtomicRMWInst::BinOp::UMax; 238 case LLVM::AtomicBinOp::umin: 239 return llvm::AtomicRMWInst::BinOp::UMin; 240 case LLVM::AtomicBinOp::fadd: 241 return llvm::AtomicRMWInst::BinOp::FAdd; 242 case LLVM::AtomicBinOp::fsub: 243 return llvm::AtomicRMWInst::BinOp::FSub; 244 } 245 llvm_unreachable("incorrect atomic binary operator"); 246 } 247 248 static llvm::AtomicOrdering getLLVMAtomicOrdering(AtomicOrdering ordering) { 249 switch (ordering) { 250 case LLVM::AtomicOrdering::not_atomic: 251 return llvm::AtomicOrdering::NotAtomic; 252 case LLVM::AtomicOrdering::unordered: 253 return llvm::AtomicOrdering::Unordered; 254 case LLVM::AtomicOrdering::monotonic: 255 return llvm::AtomicOrdering::Monotonic; 256 case LLVM::AtomicOrdering::acquire: 257 return llvm::AtomicOrdering::Acquire; 258 case LLVM::AtomicOrdering::release: 259 return llvm::AtomicOrdering::Release; 260 case LLVM::AtomicOrdering::acq_rel: 261 return llvm::AtomicOrdering::AcquireRelease; 262 case LLVM::AtomicOrdering::seq_cst: 263 return llvm::AtomicOrdering::SequentiallyConsistent; 264 } 265 llvm_unreachable("incorrect atomic ordering"); 266 } 267 268 /// Given a single MLIR operation, create the corresponding LLVM IR operation 269 /// using the `builder`. LLVM IR Builder does not have a generic interface so 270 /// this has to be a long chain of `if`s calling different functions with a 271 /// different number of arguments. 272 LogicalResult ModuleTranslation::convertOperation(Operation &opInst, 273 llvm::IRBuilder<> &builder) { 274 auto extractPosition = [](ArrayAttr attr) { 275 SmallVector<unsigned, 4> position; 276 position.reserve(attr.size()); 277 for (Attribute v : attr) 278 position.push_back(v.cast<IntegerAttr>().getValue().getZExtValue()); 279 return position; 280 }; 281 282 #include "mlir/Dialect/LLVMIR/LLVMConversions.inc" 283 284 // Emit function calls. If the "callee" attribute is present, this is a 285 // direct function call and we also need to look up the remapped function 286 // itself. Otherwise, this is an indirect call and the callee is the first 287 // operand, look it up as a normal value. Return the llvm::Value representing 288 // the function result, which may be of llvm::VoidTy type. 289 auto convertCall = [this, &builder](Operation &op) -> llvm::Value * { 290 auto operands = lookupValues(op.getOperands()); 291 ArrayRef<llvm::Value *> operandsRef(operands); 292 if (auto attr = op.getAttrOfType<FlatSymbolRefAttr>("callee")) { 293 return builder.CreateCall(functionMapping.lookup(attr.getValue()), 294 operandsRef); 295 } else { 296 return builder.CreateCall(operandsRef.front(), operandsRef.drop_front()); 297 } 298 }; 299 300 // Emit calls. If the called function has a result, remap the corresponding 301 // value. Note that LLVM IR dialect CallOp has either 0 or 1 result. 302 if (isa<LLVM::CallOp>(opInst)) { 303 llvm::Value *result = convertCall(opInst); 304 if (opInst.getNumResults() != 0) { 305 valueMapping[opInst.getResult(0)] = result; 306 return success(); 307 } 308 // Check that LLVM call returns void for 0-result functions. 309 return success(result->getType()->isVoidTy()); 310 } 311 312 if (auto invOp = dyn_cast<LLVM::InvokeOp>(opInst)) { 313 auto operands = lookupValues(opInst.getOperands()); 314 ArrayRef<llvm::Value *> operandsRef(operands); 315 if (auto attr = opInst.getAttrOfType<FlatSymbolRefAttr>("callee")) 316 builder.CreateInvoke(functionMapping.lookup(attr.getValue()), 317 blockMapping[invOp.getSuccessor(0)], 318 blockMapping[invOp.getSuccessor(1)], operandsRef); 319 else 320 builder.CreateInvoke( 321 operandsRef.front(), blockMapping[invOp.getSuccessor(0)], 322 blockMapping[invOp.getSuccessor(1)], operandsRef.drop_front()); 323 return success(); 324 } 325 326 if (auto lpOp = dyn_cast<LLVM::LandingpadOp>(opInst)) { 327 llvm::Type *ty = lpOp.getType().dyn_cast<LLVMType>().getUnderlyingType(); 328 llvm::LandingPadInst *lpi = 329 builder.CreateLandingPad(ty, lpOp.getNumOperands()); 330 331 // Add clauses 332 for (auto operand : lookupValues(lpOp.getOperands())) { 333 // All operands should be constant - checked by verifier 334 if (auto constOperand = dyn_cast<llvm::Constant>(operand)) 335 lpi->addClause(constOperand); 336 } 337 return success(); 338 } 339 340 // Emit branches. We need to look up the remapped blocks and ignore the block 341 // arguments that were transformed into PHI nodes. 342 if (auto brOp = dyn_cast<LLVM::BrOp>(opInst)) { 343 builder.CreateBr(blockMapping[brOp.getSuccessor(0)]); 344 return success(); 345 } 346 if (auto condbrOp = dyn_cast<LLVM::CondBrOp>(opInst)) { 347 builder.CreateCondBr(valueMapping.lookup(condbrOp.getOperand(0)), 348 blockMapping[condbrOp.getSuccessor(0)], 349 blockMapping[condbrOp.getSuccessor(1)]); 350 return success(); 351 } 352 353 // Emit addressof. We need to look up the global value referenced by the 354 // operation and store it in the MLIR-to-LLVM value mapping. This does not 355 // emit any LLVM instruction. 356 if (auto addressOfOp = dyn_cast<LLVM::AddressOfOp>(opInst)) { 357 LLVM::GlobalOp global = addressOfOp.getGlobal(); 358 // The verifier should not have allowed this. 359 assert(global && "referencing an undefined global"); 360 361 valueMapping[addressOfOp.getResult()] = globalsMapping.lookup(global); 362 return success(); 363 } 364 365 return opInst.emitError("unsupported or non-LLVM operation: ") 366 << opInst.getName(); 367 } 368 369 /// Convert block to LLVM IR. Unless `ignoreArguments` is set, emit PHI nodes 370 /// to define values corresponding to the MLIR block arguments. These nodes 371 /// are not connected to the source basic blocks, which may not exist yet. 372 LogicalResult ModuleTranslation::convertBlock(Block &bb, bool ignoreArguments) { 373 llvm::IRBuilder<> builder(blockMapping[&bb]); 374 375 // Before traversing operations, make block arguments available through 376 // value remapping and PHI nodes, but do not add incoming edges for the PHI 377 // nodes just yet: those values may be defined by this or following blocks. 378 // This step is omitted if "ignoreArguments" is set. The arguments of the 379 // first block have been already made available through the remapping of 380 // LLVM function arguments. 381 if (!ignoreArguments) { 382 auto predecessors = bb.getPredecessors(); 383 unsigned numPredecessors = 384 std::distance(predecessors.begin(), predecessors.end()); 385 for (auto arg : bb.getArguments()) { 386 auto wrappedType = arg.getType().dyn_cast<LLVM::LLVMType>(); 387 if (!wrappedType) 388 return emitError(bb.front().getLoc(), 389 "block argument does not have an LLVM type"); 390 llvm::Type *type = wrappedType.getUnderlyingType(); 391 llvm::PHINode *phi = builder.CreatePHI(type, numPredecessors); 392 valueMapping[arg] = phi; 393 } 394 } 395 396 // Traverse operations. 397 for (auto &op : bb) { 398 if (failed(convertOperation(op, builder))) 399 return failure(); 400 } 401 402 return success(); 403 } 404 405 /// Create named global variables that correspond to llvm.mlir.global 406 /// definitions. 407 void ModuleTranslation::convertGlobals() { 408 for (auto op : getModuleBody(mlirModule).getOps<LLVM::GlobalOp>()) { 409 llvm::Type *type = op.getType().getUnderlyingType(); 410 llvm::Constant *cst = llvm::UndefValue::get(type); 411 if (op.getValueOrNull()) { 412 // String attributes are treated separately because they cannot appear as 413 // in-function constants and are thus not supported by getLLVMConstant. 414 if (auto strAttr = op.getValueOrNull().dyn_cast_or_null<StringAttr>()) { 415 cst = llvm::ConstantDataArray::getString( 416 llvmModule->getContext(), strAttr.getValue(), /*AddNull=*/false); 417 type = cst->getType(); 418 } else { 419 cst = getLLVMConstant(type, op.getValueOrNull(), op.getLoc()); 420 } 421 } else if (Block *initializer = op.getInitializerBlock()) { 422 llvm::IRBuilder<> builder(llvmModule->getContext()); 423 for (auto &op : initializer->without_terminator()) { 424 if (failed(convertOperation(op, builder)) || 425 !isa<llvm::Constant>(valueMapping.lookup(op.getResult(0)))) { 426 emitError(op.getLoc(), "unemittable constant value"); 427 return; 428 } 429 } 430 ReturnOp ret = cast<ReturnOp>(initializer->getTerminator()); 431 cst = cast<llvm::Constant>(valueMapping.lookup(ret.getOperand(0))); 432 } 433 434 auto linkage = convertLinkageToLLVM(op.linkage()); 435 bool anyExternalLinkage = 436 (linkage == llvm::GlobalVariable::ExternalLinkage || 437 linkage == llvm::GlobalVariable::ExternalWeakLinkage); 438 auto addrSpace = op.addr_space().getLimitedValue(); 439 auto *var = new llvm::GlobalVariable( 440 *llvmModule, type, op.constant(), linkage, 441 anyExternalLinkage ? nullptr : cst, op.sym_name(), 442 /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, addrSpace); 443 444 globalsMapping.try_emplace(op, var); 445 } 446 } 447 448 /// Get the SSA value passed to the current block from the terminator operation 449 /// of its predecessor. 450 static Value getPHISourceValue(Block *current, Block *pred, 451 unsigned numArguments, unsigned index) { 452 auto &terminator = *pred->getTerminator(); 453 if (isa<LLVM::BrOp>(terminator)) { 454 return terminator.getOperand(index); 455 } 456 457 // For conditional branches, we need to check if the current block is reached 458 // through the "true" or the "false" branch and take the relevant operands. 459 auto condBranchOp = dyn_cast<LLVM::CondBrOp>(terminator); 460 assert(condBranchOp && 461 "only branch operations can be terminators of a block that " 462 "has successors"); 463 assert((condBranchOp.getSuccessor(0) != condBranchOp.getSuccessor(1)) && 464 "successors with arguments in LLVM conditional branches must be " 465 "different blocks"); 466 467 return condBranchOp.getSuccessor(0) == current 468 ? terminator.getSuccessorOperand(0, index) 469 : terminator.getSuccessorOperand(1, index); 470 } 471 472 void ModuleTranslation::connectPHINodes(LLVMFuncOp func) { 473 // Skip the first block, it cannot be branched to and its arguments correspond 474 // to the arguments of the LLVM function. 475 for (auto it = std::next(func.begin()), eit = func.end(); it != eit; ++it) { 476 Block *bb = &*it; 477 llvm::BasicBlock *llvmBB = blockMapping.lookup(bb); 478 auto phis = llvmBB->phis(); 479 auto numArguments = bb->getNumArguments(); 480 assert(numArguments == std::distance(phis.begin(), phis.end())); 481 for (auto &numberedPhiNode : llvm::enumerate(phis)) { 482 auto &phiNode = numberedPhiNode.value(); 483 unsigned index = numberedPhiNode.index(); 484 for (auto *pred : bb->getPredecessors()) { 485 phiNode.addIncoming(valueMapping.lookup(getPHISourceValue( 486 bb, pred, numArguments, index)), 487 blockMapping.lookup(pred)); 488 } 489 } 490 } 491 } 492 493 // TODO(mlir-team): implement an iterative version 494 static void topologicalSortImpl(llvm::SetVector<Block *> &blocks, Block *b) { 495 blocks.insert(b); 496 for (Block *bb : b->getSuccessors()) { 497 if (blocks.count(bb) == 0) 498 topologicalSortImpl(blocks, bb); 499 } 500 } 501 502 /// Sort function blocks topologically. 503 static llvm::SetVector<Block *> topologicalSort(LLVMFuncOp f) { 504 // For each blocks that has not been visited yet (i.e. that has no 505 // predecessors), add it to the list and traverse its successors in DFS 506 // preorder. 507 llvm::SetVector<Block *> blocks; 508 for (Block &b : f.getBlocks()) { 509 if (blocks.count(&b) == 0) 510 topologicalSortImpl(blocks, &b); 511 } 512 assert(blocks.size() == f.getBlocks().size() && "some blocks are not sorted"); 513 514 return blocks; 515 } 516 517 LogicalResult ModuleTranslation::convertOneFunction(LLVMFuncOp func) { 518 // Clear the block and value mappings, they are only relevant within one 519 // function. 520 blockMapping.clear(); 521 valueMapping.clear(); 522 llvm::Function *llvmFunc = functionMapping.lookup(func.getName()); 523 // Add function arguments to the value remapping table. 524 // If there was noalias info then we decorate each argument accordingly. 525 unsigned int argIdx = 0; 526 for (auto kvp : llvm::zip(func.getArguments(), llvmFunc->args())) { 527 llvm::Argument &llvmArg = std::get<1>(kvp); 528 BlockArgument mlirArg = std::get<0>(kvp); 529 530 if (auto attr = func.getArgAttrOfType<BoolAttr>(argIdx, "llvm.noalias")) { 531 // NB: Attribute already verified to be boolean, so check if we can indeed 532 // attach the attribute to this argument, based on its type. 533 auto argTy = mlirArg.getType().dyn_cast<LLVM::LLVMType>(); 534 if (!argTy.getUnderlyingType()->isPointerTy()) 535 return func.emitError( 536 "llvm.noalias attribute attached to LLVM non-pointer argument"); 537 if (attr.getValue()) 538 llvmArg.addAttr(llvm::Attribute::AttrKind::NoAlias); 539 } 540 valueMapping[mlirArg] = &llvmArg; 541 argIdx++; 542 } 543 544 // First, create all blocks so we can jump to them. 545 llvm::LLVMContext &llvmContext = llvmFunc->getContext(); 546 for (auto &bb : func) { 547 auto *llvmBB = llvm::BasicBlock::Create(llvmContext); 548 llvmBB->insertInto(llvmFunc); 549 blockMapping[&bb] = llvmBB; 550 } 551 552 // Then, convert blocks one by one in topological order to ensure defs are 553 // converted before uses. 554 auto blocks = topologicalSort(func); 555 for (auto indexedBB : llvm::enumerate(blocks)) { 556 auto *bb = indexedBB.value(); 557 if (failed(convertBlock(*bb, /*ignoreArguments=*/indexedBB.index() == 0))) 558 return failure(); 559 } 560 561 // Finally, after all blocks have been traversed and values mapped, connect 562 // the PHI nodes to the results of preceding blocks. 563 connectPHINodes(func); 564 return success(); 565 } 566 567 LogicalResult ModuleTranslation::checkSupportedModuleOps(Operation *m) { 568 for (Operation &o : getModuleBody(m).getOperations()) 569 if (!isa<LLVM::LLVMFuncOp>(&o) && !isa<LLVM::GlobalOp>(&o) && 570 !o.isKnownTerminator()) 571 return o.emitOpError("unsupported module-level operation"); 572 return success(); 573 } 574 575 LogicalResult ModuleTranslation::convertFunctions() { 576 // Declare all functions first because there may be function calls that form a 577 // call graph with cycles. 578 for (auto function : getModuleBody(mlirModule).getOps<LLVMFuncOp>()) { 579 llvm::FunctionCallee llvmFuncCst = llvmModule->getOrInsertFunction( 580 function.getName(), 581 cast<llvm::FunctionType>(function.getType().getUnderlyingType())); 582 assert(isa<llvm::Function>(llvmFuncCst.getCallee())); 583 functionMapping[function.getName()] = 584 cast<llvm::Function>(llvmFuncCst.getCallee()); 585 } 586 587 // Convert functions. 588 for (auto function : getModuleBody(mlirModule).getOps<LLVMFuncOp>()) { 589 // Ignore external functions. 590 if (function.isExternal()) 591 continue; 592 593 if (failed(convertOneFunction(function))) 594 return failure(); 595 } 596 597 return success(); 598 } 599 600 /// A helper to look up remapped operands in the value remapping table.` 601 SmallVector<llvm::Value *, 8> 602 ModuleTranslation::lookupValues(ValueRange values) { 603 SmallVector<llvm::Value *, 8> remapped; 604 remapped.reserve(values.size()); 605 for (Value v : values) 606 remapped.push_back(valueMapping.lookup(v)); 607 return remapped; 608 } 609 610 std::unique_ptr<llvm::Module> 611 ModuleTranslation::prepareLLVMModule(Operation *m) { 612 auto *dialect = m->getContext()->getRegisteredDialect<LLVM::LLVMDialect>(); 613 assert(dialect && "LLVM dialect must be registered"); 614 615 auto llvmModule = llvm::CloneModule(dialect->getLLVMModule()); 616 if (!llvmModule) 617 return nullptr; 618 619 llvm::LLVMContext &llvmContext = llvmModule->getContext(); 620 llvm::IRBuilder<> builder(llvmContext); 621 622 // Inject declarations for `malloc` and `free` functions that can be used in 623 // memref allocation/deallocation coming from standard ops lowering. 624 llvmModule->getOrInsertFunction("malloc", builder.getInt8PtrTy(), 625 builder.getInt64Ty()); 626 llvmModule->getOrInsertFunction("free", builder.getVoidTy(), 627 builder.getInt8PtrTy()); 628 629 return llvmModule; 630 } 631