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 "DebugTranslation.h"
17 #include "mlir/ADT/TypeSwitch.h"
18 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
19 #include "mlir/Dialect/OpenMP/OpenMPDialect.h"
20 #include "mlir/IR/Attributes.h"
21 #include "mlir/IR/Module.h"
22 #include "mlir/IR/StandardTypes.h"
23 #include "mlir/Support/LLVM.h"
24 
25 #include "llvm/ADT/SetVector.h"
26 #include "llvm/Frontend/OpenMP/OMPIRBuilder.h"
27 #include "llvm/IR/BasicBlock.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/LLVMContext.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/Transforms/Utils/Cloning.h"
34 
35 using namespace mlir;
36 using namespace mlir::LLVM;
37 using namespace mlir::LLVM::detail;
38 
39 #include "mlir/Dialect/LLVMIR/LLVMConversionEnumsToLLVM.inc"
40 
41 /// Builds a constant of a sequential LLVM type `type`, potentially containing
42 /// other sequential types recursively, from the individual constant values
43 /// provided in `constants`. `shape` contains the number of elements in nested
44 /// sequential types. Reports errors at `loc` and returns nullptr on error.
45 static llvm::Constant *
46 buildSequentialConstant(ArrayRef<llvm::Constant *> &constants,
47                         ArrayRef<int64_t> shape, llvm::Type *type,
48                         Location loc) {
49   if (shape.empty()) {
50     llvm::Constant *result = constants.front();
51     constants = constants.drop_front();
52     return result;
53   }
54 
55   llvm::Type *elementType;
56   if (auto *arrayTy = dyn_cast<llvm::ArrayType>(type)) {
57     elementType = arrayTy->getElementType();
58   } else if (auto *vectorTy = dyn_cast<llvm::VectorType>(type)) {
59     elementType = vectorTy->getElementType();
60   } else {
61     emitError(loc) << "expected sequential LLVM types wrapping a scalar";
62     return nullptr;
63   }
64 
65   SmallVector<llvm::Constant *, 8> nested;
66   nested.reserve(shape.front());
67   for (int64_t i = 0; i < shape.front(); ++i) {
68     nested.push_back(buildSequentialConstant(constants, shape.drop_front(),
69                                              elementType, loc));
70     if (!nested.back())
71       return nullptr;
72   }
73 
74   if (shape.size() == 1 && type->isVectorTy())
75     return llvm::ConstantVector::get(nested);
76   return llvm::ConstantArray::get(
77       llvm::ArrayType::get(elementType, shape.front()), nested);
78 }
79 
80 /// Returns the first non-sequential type nested in sequential types.
81 static llvm::Type *getInnermostElementType(llvm::Type *type) {
82   do {
83     if (auto *arrayTy = dyn_cast<llvm::ArrayType>(type)) {
84       type = arrayTy->getElementType();
85     } else if (auto *vectorTy = dyn_cast<llvm::VectorType>(type)) {
86       type = vectorTy->getElementType();
87     } else {
88       return type;
89     }
90   } while (1);
91 }
92 
93 /// Create an LLVM IR constant of `llvmType` from the MLIR attribute `attr`.
94 /// This currently supports integer, floating point, splat and dense element
95 /// attributes and combinations thereof.  In case of error, report it to `loc`
96 /// and return nullptr.
97 llvm::Constant *ModuleTranslation::getLLVMConstant(llvm::Type *llvmType,
98                                                    Attribute attr,
99                                                    Location loc) {
100   if (!attr)
101     return llvm::UndefValue::get(llvmType);
102   if (llvmType->isStructTy()) {
103     emitError(loc, "struct types are not supported in constants");
104     return nullptr;
105   }
106   // For integer types, we allow a mismatch in sizes as the index type in
107   // MLIR might have a different size than the index type in the LLVM module.
108   if (auto intAttr = attr.dyn_cast<IntegerAttr>())
109     return llvm::ConstantInt::get(
110         llvmType,
111         intAttr.getValue().sextOrTrunc(llvmType->getIntegerBitWidth()));
112   if (auto boolAttr = attr.dyn_cast<BoolAttr>())
113     return llvm::ConstantInt::get(llvmType, boolAttr.getValue());
114   if (auto floatAttr = attr.dyn_cast<FloatAttr>())
115     return llvm::ConstantFP::get(llvmType, floatAttr.getValue());
116   if (auto funcAttr = attr.dyn_cast<FlatSymbolRefAttr>())
117     return llvm::ConstantExpr::getBitCast(
118         functionMapping.lookup(funcAttr.getValue()), llvmType);
119   if (auto splatAttr = attr.dyn_cast<SplatElementsAttr>()) {
120     llvm::Type *elementType;
121     uint64_t numElements;
122     if (auto *arrayTy = dyn_cast<llvm::ArrayType>(llvmType)) {
123       elementType = arrayTy->getElementType();
124       numElements = arrayTy->getNumElements();
125     } else {
126       auto *vectorTy = cast<llvm::VectorType>(llvmType);
127       elementType = vectorTy->getElementType();
128       numElements = vectorTy->getNumElements();
129     }
130     // Splat value is a scalar. Extract it only if the element type is not
131     // another sequence type. The recursion terminates because each step removes
132     // one outer sequential type.
133     bool elementTypeSequential =
134         isa<llvm::ArrayType>(elementType) || isa<llvm::VectorType>(elementType);
135     llvm::Constant *child = getLLVMConstant(
136         elementType,
137         elementTypeSequential ? splatAttr : splatAttr.getSplatValue(), loc);
138     if (!child)
139       return nullptr;
140     if (llvmType->isVectorTy())
141       return llvm::ConstantVector::getSplat(
142           llvm::ElementCount(numElements, /*Scalable=*/false), child);
143     if (llvmType->isArrayTy()) {
144       auto *arrayType = llvm::ArrayType::get(elementType, numElements);
145       SmallVector<llvm::Constant *, 8> constants(numElements, child);
146       return llvm::ConstantArray::get(arrayType, constants);
147     }
148   }
149 
150   if (auto elementsAttr = attr.dyn_cast<ElementsAttr>()) {
151     assert(elementsAttr.getType().hasStaticShape());
152     assert(elementsAttr.getNumElements() != 0 &&
153            "unexpected empty elements attribute");
154     assert(!elementsAttr.getType().getShape().empty() &&
155            "unexpected empty elements attribute shape");
156 
157     SmallVector<llvm::Constant *, 8> constants;
158     constants.reserve(elementsAttr.getNumElements());
159     llvm::Type *innermostType = getInnermostElementType(llvmType);
160     for (auto n : elementsAttr.getValues<Attribute>()) {
161       constants.push_back(getLLVMConstant(innermostType, n, loc));
162       if (!constants.back())
163         return nullptr;
164     }
165     ArrayRef<llvm::Constant *> constantsRef = constants;
166     llvm::Constant *result = buildSequentialConstant(
167         constantsRef, elementsAttr.getType().getShape(), llvmType, loc);
168     assert(constantsRef.empty() && "did not consume all elemental constants");
169     return result;
170   }
171 
172   if (auto stringAttr = attr.dyn_cast<StringAttr>()) {
173     return llvm::ConstantDataArray::get(
174         llvmModule->getContext(), ArrayRef<char>{stringAttr.getValue().data(),
175                                                  stringAttr.getValue().size()});
176   }
177   emitError(loc, "unsupported constant value");
178   return nullptr;
179 }
180 
181 /// Convert MLIR integer comparison predicate to LLVM IR comparison predicate.
182 static llvm::CmpInst::Predicate getLLVMCmpPredicate(ICmpPredicate p) {
183   switch (p) {
184   case LLVM::ICmpPredicate::eq:
185     return llvm::CmpInst::Predicate::ICMP_EQ;
186   case LLVM::ICmpPredicate::ne:
187     return llvm::CmpInst::Predicate::ICMP_NE;
188   case LLVM::ICmpPredicate::slt:
189     return llvm::CmpInst::Predicate::ICMP_SLT;
190   case LLVM::ICmpPredicate::sle:
191     return llvm::CmpInst::Predicate::ICMP_SLE;
192   case LLVM::ICmpPredicate::sgt:
193     return llvm::CmpInst::Predicate::ICMP_SGT;
194   case LLVM::ICmpPredicate::sge:
195     return llvm::CmpInst::Predicate::ICMP_SGE;
196   case LLVM::ICmpPredicate::ult:
197     return llvm::CmpInst::Predicate::ICMP_ULT;
198   case LLVM::ICmpPredicate::ule:
199     return llvm::CmpInst::Predicate::ICMP_ULE;
200   case LLVM::ICmpPredicate::ugt:
201     return llvm::CmpInst::Predicate::ICMP_UGT;
202   case LLVM::ICmpPredicate::uge:
203     return llvm::CmpInst::Predicate::ICMP_UGE;
204   }
205   llvm_unreachable("incorrect comparison predicate");
206 }
207 
208 static llvm::CmpInst::Predicate getLLVMCmpPredicate(FCmpPredicate p) {
209   switch (p) {
210   case LLVM::FCmpPredicate::_false:
211     return llvm::CmpInst::Predicate::FCMP_FALSE;
212   case LLVM::FCmpPredicate::oeq:
213     return llvm::CmpInst::Predicate::FCMP_OEQ;
214   case LLVM::FCmpPredicate::ogt:
215     return llvm::CmpInst::Predicate::FCMP_OGT;
216   case LLVM::FCmpPredicate::oge:
217     return llvm::CmpInst::Predicate::FCMP_OGE;
218   case LLVM::FCmpPredicate::olt:
219     return llvm::CmpInst::Predicate::FCMP_OLT;
220   case LLVM::FCmpPredicate::ole:
221     return llvm::CmpInst::Predicate::FCMP_OLE;
222   case LLVM::FCmpPredicate::one:
223     return llvm::CmpInst::Predicate::FCMP_ONE;
224   case LLVM::FCmpPredicate::ord:
225     return llvm::CmpInst::Predicate::FCMP_ORD;
226   case LLVM::FCmpPredicate::ueq:
227     return llvm::CmpInst::Predicate::FCMP_UEQ;
228   case LLVM::FCmpPredicate::ugt:
229     return llvm::CmpInst::Predicate::FCMP_UGT;
230   case LLVM::FCmpPredicate::uge:
231     return llvm::CmpInst::Predicate::FCMP_UGE;
232   case LLVM::FCmpPredicate::ult:
233     return llvm::CmpInst::Predicate::FCMP_ULT;
234   case LLVM::FCmpPredicate::ule:
235     return llvm::CmpInst::Predicate::FCMP_ULE;
236   case LLVM::FCmpPredicate::une:
237     return llvm::CmpInst::Predicate::FCMP_UNE;
238   case LLVM::FCmpPredicate::uno:
239     return llvm::CmpInst::Predicate::FCMP_UNO;
240   case LLVM::FCmpPredicate::_true:
241     return llvm::CmpInst::Predicate::FCMP_TRUE;
242   }
243   llvm_unreachable("incorrect comparison predicate");
244 }
245 
246 static llvm::AtomicRMWInst::BinOp getLLVMAtomicBinOp(AtomicBinOp op) {
247   switch (op) {
248   case LLVM::AtomicBinOp::xchg:
249     return llvm::AtomicRMWInst::BinOp::Xchg;
250   case LLVM::AtomicBinOp::add:
251     return llvm::AtomicRMWInst::BinOp::Add;
252   case LLVM::AtomicBinOp::sub:
253     return llvm::AtomicRMWInst::BinOp::Sub;
254   case LLVM::AtomicBinOp::_and:
255     return llvm::AtomicRMWInst::BinOp::And;
256   case LLVM::AtomicBinOp::nand:
257     return llvm::AtomicRMWInst::BinOp::Nand;
258   case LLVM::AtomicBinOp::_or:
259     return llvm::AtomicRMWInst::BinOp::Or;
260   case LLVM::AtomicBinOp::_xor:
261     return llvm::AtomicRMWInst::BinOp::Xor;
262   case LLVM::AtomicBinOp::max:
263     return llvm::AtomicRMWInst::BinOp::Max;
264   case LLVM::AtomicBinOp::min:
265     return llvm::AtomicRMWInst::BinOp::Min;
266   case LLVM::AtomicBinOp::umax:
267     return llvm::AtomicRMWInst::BinOp::UMax;
268   case LLVM::AtomicBinOp::umin:
269     return llvm::AtomicRMWInst::BinOp::UMin;
270   case LLVM::AtomicBinOp::fadd:
271     return llvm::AtomicRMWInst::BinOp::FAdd;
272   case LLVM::AtomicBinOp::fsub:
273     return llvm::AtomicRMWInst::BinOp::FSub;
274   }
275   llvm_unreachable("incorrect atomic binary operator");
276 }
277 
278 static llvm::AtomicOrdering getLLVMAtomicOrdering(AtomicOrdering ordering) {
279   switch (ordering) {
280   case LLVM::AtomicOrdering::not_atomic:
281     return llvm::AtomicOrdering::NotAtomic;
282   case LLVM::AtomicOrdering::unordered:
283     return llvm::AtomicOrdering::Unordered;
284   case LLVM::AtomicOrdering::monotonic:
285     return llvm::AtomicOrdering::Monotonic;
286   case LLVM::AtomicOrdering::acquire:
287     return llvm::AtomicOrdering::Acquire;
288   case LLVM::AtomicOrdering::release:
289     return llvm::AtomicOrdering::Release;
290   case LLVM::AtomicOrdering::acq_rel:
291     return llvm::AtomicOrdering::AcquireRelease;
292   case LLVM::AtomicOrdering::seq_cst:
293     return llvm::AtomicOrdering::SequentiallyConsistent;
294   }
295   llvm_unreachable("incorrect atomic ordering");
296 }
297 
298 ModuleTranslation::ModuleTranslation(Operation *module,
299                                      std::unique_ptr<llvm::Module> llvmModule)
300     : mlirModule(module), llvmModule(std::move(llvmModule)),
301       debugTranslation(
302           std::make_unique<DebugTranslation>(module, *this->llvmModule)),
303       ompDialect(
304           module->getContext()->getRegisteredDialect<omp::OpenMPDialect>()) {
305   assert(satisfiesLLVMModule(mlirModule) &&
306          "mlirModule should honor LLVM's module semantics.");
307 }
308 ModuleTranslation::~ModuleTranslation() {}
309 
310 /// Given an OpenMP MLIR operation, create the corresponding LLVM IR
311 /// (including OpenMP runtime calls).
312 LogicalResult
313 ModuleTranslation::convertOmpOperation(Operation &opInst,
314                                        llvm::IRBuilder<> &builder) {
315   if (!ompBuilder) {
316     ompBuilder = std::make_unique<llvm::OpenMPIRBuilder>(*llvmModule);
317     ompBuilder->initialize();
318   }
319   return mlir::TypeSwitch<Operation *, LogicalResult>(&opInst)
320       .Case([&](omp::BarrierOp) {
321         ompBuilder->CreateBarrier(builder.saveIP(), llvm::omp::OMPD_barrier);
322         return success();
323       })
324       .Case([&](omp::TaskwaitOp) {
325         ompBuilder->CreateTaskwait(builder.saveIP());
326         return success();
327       })
328       .Case([&](omp::TaskyieldOp) {
329         ompBuilder->CreateTaskyield(builder.saveIP());
330         return success();
331       })
332       .Default([&](Operation *inst) {
333         return inst->emitError("unsupported OpenMP operation: ")
334                << inst->getName();
335       });
336 }
337 
338 /// Given a single MLIR operation, create the corresponding LLVM IR operation
339 /// using the `builder`.  LLVM IR Builder does not have a generic interface so
340 /// this has to be a long chain of `if`s calling different functions with a
341 /// different number of arguments.
342 LogicalResult ModuleTranslation::convertOperation(Operation &opInst,
343                                                   llvm::IRBuilder<> &builder) {
344   auto extractPosition = [](ArrayAttr attr) {
345     SmallVector<unsigned, 4> position;
346     position.reserve(attr.size());
347     for (Attribute v : attr)
348       position.push_back(v.cast<IntegerAttr>().getValue().getZExtValue());
349     return position;
350   };
351 
352 #include "mlir/Dialect/LLVMIR/LLVMConversions.inc"
353 
354   // Emit function calls.  If the "callee" attribute is present, this is a
355   // direct function call and we also need to look up the remapped function
356   // itself.  Otherwise, this is an indirect call and the callee is the first
357   // operand, look it up as a normal value.  Return the llvm::Value representing
358   // the function result, which may be of llvm::VoidTy type.
359   auto convertCall = [this, &builder](Operation &op) -> llvm::Value * {
360     auto operands = lookupValues(op.getOperands());
361     ArrayRef<llvm::Value *> operandsRef(operands);
362     if (auto attr = op.getAttrOfType<FlatSymbolRefAttr>("callee")) {
363       return builder.CreateCall(functionMapping.lookup(attr.getValue()),
364                                 operandsRef);
365     } else {
366       auto *calleePtrType =
367           cast<llvm::PointerType>(operandsRef.front()->getType());
368       auto *calleeType =
369           cast<llvm::FunctionType>(calleePtrType->getElementType());
370       return builder.CreateCall(calleeType, operandsRef.front(),
371                                 operandsRef.drop_front());
372     }
373   };
374 
375   // Emit calls.  If the called function has a result, remap the corresponding
376   // value.  Note that LLVM IR dialect CallOp has either 0 or 1 result.
377   if (isa<LLVM::CallOp>(opInst)) {
378     llvm::Value *result = convertCall(opInst);
379     if (opInst.getNumResults() != 0) {
380       valueMapping[opInst.getResult(0)] = result;
381       return success();
382     }
383     // Check that LLVM call returns void for 0-result functions.
384     return success(result->getType()->isVoidTy());
385   }
386 
387   if (auto invOp = dyn_cast<LLVM::InvokeOp>(opInst)) {
388     auto operands = lookupValues(opInst.getOperands());
389     ArrayRef<llvm::Value *> operandsRef(operands);
390     if (auto attr = opInst.getAttrOfType<FlatSymbolRefAttr>("callee")) {
391       builder.CreateInvoke(functionMapping.lookup(attr.getValue()),
392                            blockMapping[invOp.getSuccessor(0)],
393                            blockMapping[invOp.getSuccessor(1)], operandsRef);
394     } else {
395       auto *calleePtrType =
396           cast<llvm::PointerType>(operandsRef.front()->getType());
397       auto *calleeType =
398           cast<llvm::FunctionType>(calleePtrType->getElementType());
399       builder.CreateInvoke(
400           calleeType, operandsRef.front(), blockMapping[invOp.getSuccessor(0)],
401           blockMapping[invOp.getSuccessor(1)], operandsRef.drop_front());
402     }
403     return success();
404   }
405 
406   if (auto lpOp = dyn_cast<LLVM::LandingpadOp>(opInst)) {
407     llvm::Type *ty = lpOp.getType().dyn_cast<LLVMType>().getUnderlyingType();
408     llvm::LandingPadInst *lpi =
409         builder.CreateLandingPad(ty, lpOp.getNumOperands());
410 
411     // Add clauses
412     for (auto operand : lookupValues(lpOp.getOperands())) {
413       // All operands should be constant - checked by verifier
414       if (auto constOperand = dyn_cast<llvm::Constant>(operand))
415         lpi->addClause(constOperand);
416     }
417     valueMapping[lpOp.getResult()] = lpi;
418     return success();
419   }
420 
421   // Emit branches.  We need to look up the remapped blocks and ignore the block
422   // arguments that were transformed into PHI nodes.
423   if (auto brOp = dyn_cast<LLVM::BrOp>(opInst)) {
424     builder.CreateBr(blockMapping[brOp.getSuccessor()]);
425     return success();
426   }
427   if (auto condbrOp = dyn_cast<LLVM::CondBrOp>(opInst)) {
428     builder.CreateCondBr(valueMapping.lookup(condbrOp.getOperand(0)),
429                          blockMapping[condbrOp.getSuccessor(0)],
430                          blockMapping[condbrOp.getSuccessor(1)]);
431     return success();
432   }
433 
434   // Emit addressof.  We need to look up the global value referenced by the
435   // operation and store it in the MLIR-to-LLVM value mapping.  This does not
436   // emit any LLVM instruction.
437   if (auto addressOfOp = dyn_cast<LLVM::AddressOfOp>(opInst)) {
438     LLVM::GlobalOp global = addressOfOp.getGlobal();
439     // The verifier should not have allowed this.
440     assert(global && "referencing an undefined global");
441 
442     valueMapping[addressOfOp.getResult()] = globalsMapping.lookup(global);
443     return success();
444   }
445 
446   if (opInst.getDialect() == ompDialect) {
447     return convertOmpOperation(opInst, builder);
448   }
449 
450   return opInst.emitError("unsupported or non-LLVM operation: ")
451          << opInst.getName();
452 }
453 
454 /// Convert block to LLVM IR.  Unless `ignoreArguments` is set, emit PHI nodes
455 /// to define values corresponding to the MLIR block arguments.  These nodes
456 /// are not connected to the source basic blocks, which may not exist yet.
457 LogicalResult ModuleTranslation::convertBlock(Block &bb, bool ignoreArguments) {
458   llvm::IRBuilder<> builder(blockMapping[&bb]);
459   auto *subprogram = builder.GetInsertBlock()->getParent()->getSubprogram();
460 
461   // Before traversing operations, make block arguments available through
462   // value remapping and PHI nodes, but do not add incoming edges for the PHI
463   // nodes just yet: those values may be defined by this or following blocks.
464   // This step is omitted if "ignoreArguments" is set.  The arguments of the
465   // first block have been already made available through the remapping of
466   // LLVM function arguments.
467   if (!ignoreArguments) {
468     auto predecessors = bb.getPredecessors();
469     unsigned numPredecessors =
470         std::distance(predecessors.begin(), predecessors.end());
471     for (auto arg : bb.getArguments()) {
472       auto wrappedType = arg.getType().dyn_cast<LLVM::LLVMType>();
473       if (!wrappedType)
474         return emitError(bb.front().getLoc(),
475                          "block argument does not have an LLVM type");
476       llvm::Type *type = wrappedType.getUnderlyingType();
477       llvm::PHINode *phi = builder.CreatePHI(type, numPredecessors);
478       valueMapping[arg] = phi;
479     }
480   }
481 
482   // Traverse operations.
483   for (auto &op : bb) {
484     // Set the current debug location within the builder.
485     builder.SetCurrentDebugLocation(
486         debugTranslation->translateLoc(op.getLoc(), subprogram));
487 
488     if (failed(convertOperation(op, builder)))
489       return failure();
490   }
491 
492   return success();
493 }
494 
495 /// Create named global variables that correspond to llvm.mlir.global
496 /// definitions.
497 LogicalResult ModuleTranslation::convertGlobals() {
498   for (auto op : getModuleBody(mlirModule).getOps<LLVM::GlobalOp>()) {
499     llvm::Type *type = op.getType().getUnderlyingType();
500     llvm::Constant *cst = llvm::UndefValue::get(type);
501     if (op.getValueOrNull()) {
502       // String attributes are treated separately because they cannot appear as
503       // in-function constants and are thus not supported by getLLVMConstant.
504       if (auto strAttr = op.getValueOrNull().dyn_cast_or_null<StringAttr>()) {
505         cst = llvm::ConstantDataArray::getString(
506             llvmModule->getContext(), strAttr.getValue(), /*AddNull=*/false);
507         type = cst->getType();
508       } else if (!(cst = getLLVMConstant(type, op.getValueOrNull(),
509                                          op.getLoc()))) {
510         return failure();
511       }
512     } else if (Block *initializer = op.getInitializerBlock()) {
513       llvm::IRBuilder<> builder(llvmModule->getContext());
514       for (auto &op : initializer->without_terminator()) {
515         if (failed(convertOperation(op, builder)) ||
516             !isa<llvm::Constant>(valueMapping.lookup(op.getResult(0))))
517           return emitError(op.getLoc(), "unemittable constant value");
518       }
519       ReturnOp ret = cast<ReturnOp>(initializer->getTerminator());
520       cst = cast<llvm::Constant>(valueMapping.lookup(ret.getOperand(0)));
521     }
522 
523     auto linkage = convertLinkageToLLVM(op.linkage());
524     bool anyExternalLinkage =
525         ((linkage == llvm::GlobalVariable::ExternalLinkage &&
526           isa<llvm::UndefValue>(cst)) ||
527          linkage == llvm::GlobalVariable::ExternalWeakLinkage);
528     auto addrSpace = op.addr_space().getLimitedValue();
529     auto *var = new llvm::GlobalVariable(
530         *llvmModule, type, op.constant(), linkage,
531         anyExternalLinkage ? nullptr : cst, op.sym_name(),
532         /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, addrSpace);
533 
534     globalsMapping.try_emplace(op, var);
535   }
536 
537   return success();
538 }
539 
540 /// Get the SSA value passed to the current block from the terminator operation
541 /// of its predecessor.
542 static Value getPHISourceValue(Block *current, Block *pred,
543                                unsigned numArguments, unsigned index) {
544   auto &terminator = *pred->getTerminator();
545   if (isa<LLVM::BrOp>(terminator)) {
546     return terminator.getOperand(index);
547   }
548 
549   // For conditional branches, we need to check if the current block is reached
550   // through the "true" or the "false" branch and take the relevant operands.
551   auto condBranchOp = dyn_cast<LLVM::CondBrOp>(terminator);
552   assert(condBranchOp &&
553          "only branch operations can be terminators of a block that "
554          "has successors");
555   assert((condBranchOp.getSuccessor(0) != condBranchOp.getSuccessor(1)) &&
556          "successors with arguments in LLVM conditional branches must be "
557          "different blocks");
558 
559   return condBranchOp.getSuccessor(0) == current
560              ? condBranchOp.trueDestOperands()[index]
561              : condBranchOp.falseDestOperands()[index];
562 }
563 
564 void ModuleTranslation::connectPHINodes(LLVMFuncOp func) {
565   // Skip the first block, it cannot be branched to and its arguments correspond
566   // to the arguments of the LLVM function.
567   for (auto it = std::next(func.begin()), eit = func.end(); it != eit; ++it) {
568     Block *bb = &*it;
569     llvm::BasicBlock *llvmBB = blockMapping.lookup(bb);
570     auto phis = llvmBB->phis();
571     auto numArguments = bb->getNumArguments();
572     assert(numArguments == std::distance(phis.begin(), phis.end()));
573     for (auto &numberedPhiNode : llvm::enumerate(phis)) {
574       auto &phiNode = numberedPhiNode.value();
575       unsigned index = numberedPhiNode.index();
576       for (auto *pred : bb->getPredecessors()) {
577         phiNode.addIncoming(valueMapping.lookup(getPHISourceValue(
578                                 bb, pred, numArguments, index)),
579                             blockMapping.lookup(pred));
580       }
581     }
582   }
583 }
584 
585 // TODO(mlir-team): implement an iterative version
586 static void topologicalSortImpl(llvm::SetVector<Block *> &blocks, Block *b) {
587   blocks.insert(b);
588   for (Block *bb : b->getSuccessors()) {
589     if (blocks.count(bb) == 0)
590       topologicalSortImpl(blocks, bb);
591   }
592 }
593 
594 /// Sort function blocks topologically.
595 static llvm::SetVector<Block *> topologicalSort(LLVMFuncOp f) {
596   // For each blocks that has not been visited yet (i.e. that has no
597   // predecessors), add it to the list and traverse its successors in DFS
598   // preorder.
599   llvm::SetVector<Block *> blocks;
600   for (Block &b : f.getBlocks()) {
601     if (blocks.count(&b) == 0)
602       topologicalSortImpl(blocks, &b);
603   }
604   assert(blocks.size() == f.getBlocks().size() && "some blocks are not sorted");
605 
606   return blocks;
607 }
608 
609 /// Attempts to add an attribute identified by `key`, optionally with the given
610 /// `value` to LLVM function `llvmFunc`. Reports errors at `loc` if any. If the
611 /// attribute has a kind known to LLVM IR, create the attribute of this kind,
612 /// otherwise keep it as a string attribute. Performs additional checks for
613 /// attributes known to have or not have a value in order to avoid assertions
614 /// inside LLVM upon construction.
615 static LogicalResult checkedAddLLVMFnAttribute(Location loc,
616                                                llvm::Function *llvmFunc,
617                                                StringRef key,
618                                                StringRef value = StringRef()) {
619   auto kind = llvm::Attribute::getAttrKindFromName(key);
620   if (kind == llvm::Attribute::None) {
621     llvmFunc->addFnAttr(key, value);
622     return success();
623   }
624 
625   if (llvm::Attribute::doesAttrKindHaveArgument(kind)) {
626     if (value.empty())
627       return emitError(loc) << "LLVM attribute '" << key << "' expects a value";
628 
629     int result;
630     if (!value.getAsInteger(/*Radix=*/0, result))
631       llvmFunc->addFnAttr(
632           llvm::Attribute::get(llvmFunc->getContext(), kind, result));
633     else
634       llvmFunc->addFnAttr(key, value);
635     return success();
636   }
637 
638   if (!value.empty())
639     return emitError(loc) << "LLVM attribute '" << key
640                           << "' does not expect a value, found '" << value
641                           << "'";
642 
643   llvmFunc->addFnAttr(kind);
644   return success();
645 }
646 
647 /// Attaches the attributes listed in the given array attribute to `llvmFunc`.
648 /// Reports error to `loc` if any and returns immediately. Expects `attributes`
649 /// to be an array attribute containing either string attributes, treated as
650 /// value-less LLVM attributes, or array attributes containing two string
651 /// attributes, with the first string being the name of the corresponding LLVM
652 /// attribute and the second string beings its value. Note that even integer
653 /// attributes are expected to have their values expressed as strings.
654 static LogicalResult
655 forwardPassthroughAttributes(Location loc, Optional<ArrayAttr> attributes,
656                              llvm::Function *llvmFunc) {
657   if (!attributes)
658     return success();
659 
660   for (Attribute attr : *attributes) {
661     if (auto stringAttr = attr.dyn_cast<StringAttr>()) {
662       if (failed(
663               checkedAddLLVMFnAttribute(loc, llvmFunc, stringAttr.getValue())))
664         return failure();
665       continue;
666     }
667 
668     auto arrayAttr = attr.dyn_cast<ArrayAttr>();
669     if (!arrayAttr || arrayAttr.size() != 2)
670       return emitError(loc)
671              << "expected 'passthrough' to contain string or array attributes";
672 
673     auto keyAttr = arrayAttr[0].dyn_cast<StringAttr>();
674     auto valueAttr = arrayAttr[1].dyn_cast<StringAttr>();
675     if (!keyAttr || !valueAttr)
676       return emitError(loc)
677              << "expected arrays within 'passthrough' to contain two strings";
678 
679     if (failed(checkedAddLLVMFnAttribute(loc, llvmFunc, keyAttr.getValue(),
680                                          valueAttr.getValue())))
681       return failure();
682   }
683   return success();
684 }
685 
686 LogicalResult ModuleTranslation::convertOneFunction(LLVMFuncOp func) {
687   // Clear the block and value mappings, they are only relevant within one
688   // function.
689   blockMapping.clear();
690   valueMapping.clear();
691   llvm::Function *llvmFunc = functionMapping.lookup(func.getName());
692 
693   // Translate the debug information for this function.
694   debugTranslation->translate(func, *llvmFunc);
695 
696   // Add function arguments to the value remapping table.
697   // If there was noalias info then we decorate each argument accordingly.
698   unsigned int argIdx = 0;
699   for (auto kvp : llvm::zip(func.getArguments(), llvmFunc->args())) {
700     llvm::Argument &llvmArg = std::get<1>(kvp);
701     BlockArgument mlirArg = std::get<0>(kvp);
702 
703     if (auto attr = func.getArgAttrOfType<BoolAttr>(argIdx, "llvm.noalias")) {
704       // NB: Attribute already verified to be boolean, so check if we can indeed
705       // attach the attribute to this argument, based on its type.
706       auto argTy = mlirArg.getType().dyn_cast<LLVM::LLVMType>();
707       if (!argTy.getUnderlyingType()->isPointerTy())
708         return func.emitError(
709             "llvm.noalias attribute attached to LLVM non-pointer argument");
710       if (attr.getValue())
711         llvmArg.addAttr(llvm::Attribute::AttrKind::NoAlias);
712     }
713     valueMapping[mlirArg] = &llvmArg;
714     argIdx++;
715   }
716 
717   // Check the personality and set it.
718   if (func.personality().hasValue()) {
719     llvm::Type *ty = llvm::Type::getInt8PtrTy(llvmFunc->getContext());
720     if (llvm::Constant *pfunc =
721             getLLVMConstant(ty, func.personalityAttr(), func.getLoc()))
722       llvmFunc->setPersonalityFn(pfunc);
723   }
724 
725   // First, create all blocks so we can jump to them.
726   llvm::LLVMContext &llvmContext = llvmFunc->getContext();
727   for (auto &bb : func) {
728     auto *llvmBB = llvm::BasicBlock::Create(llvmContext);
729     llvmBB->insertInto(llvmFunc);
730     blockMapping[&bb] = llvmBB;
731   }
732 
733   // Then, convert blocks one by one in topological order to ensure defs are
734   // converted before uses.
735   auto blocks = topologicalSort(func);
736   for (auto indexedBB : llvm::enumerate(blocks)) {
737     auto *bb = indexedBB.value();
738     if (failed(convertBlock(*bb, /*ignoreArguments=*/indexedBB.index() == 0)))
739       return failure();
740   }
741 
742   // Finally, after all blocks have been traversed and values mapped, connect
743   // the PHI nodes to the results of preceding blocks.
744   connectPHINodes(func);
745   return success();
746 }
747 
748 LogicalResult ModuleTranslation::checkSupportedModuleOps(Operation *m) {
749   for (Operation &o : getModuleBody(m).getOperations())
750     if (!isa<LLVM::LLVMFuncOp>(&o) && !isa<LLVM::GlobalOp>(&o) &&
751         !o.isKnownTerminator())
752       return o.emitOpError("unsupported module-level operation");
753   return success();
754 }
755 
756 LogicalResult ModuleTranslation::convertFunctions() {
757   // Declare all functions first because there may be function calls that form a
758   // call graph with cycles.
759   for (auto function : getModuleBody(mlirModule).getOps<LLVMFuncOp>()) {
760     llvm::FunctionCallee llvmFuncCst = llvmModule->getOrInsertFunction(
761         function.getName(),
762         cast<llvm::FunctionType>(function.getType().getUnderlyingType()));
763     llvm::Function *llvmFunc = cast<llvm::Function>(llvmFuncCst.getCallee());
764     functionMapping[function.getName()] = llvmFunc;
765 
766     // Forward the pass-through attributes to LLVM.
767     if (failed(forwardPassthroughAttributes(function.getLoc(),
768                                             function.passthrough(), llvmFunc)))
769       return failure();
770   }
771 
772   // Convert functions.
773   for (auto function : getModuleBody(mlirModule).getOps<LLVMFuncOp>()) {
774     // Ignore external functions.
775     if (function.isExternal())
776       continue;
777 
778     if (failed(convertOneFunction(function)))
779       return failure();
780   }
781 
782   return success();
783 }
784 
785 /// A helper to look up remapped operands in the value remapping table.`
786 SmallVector<llvm::Value *, 8>
787 ModuleTranslation::lookupValues(ValueRange values) {
788   SmallVector<llvm::Value *, 8> remapped;
789   remapped.reserve(values.size());
790   for (Value v : values) {
791     assert(valueMapping.count(v) && "referencing undefined value");
792     remapped.push_back(valueMapping.lookup(v));
793   }
794   return remapped;
795 }
796 
797 std::unique_ptr<llvm::Module>
798 ModuleTranslation::prepareLLVMModule(Operation *m) {
799   auto *dialect = m->getContext()->getRegisteredDialect<LLVM::LLVMDialect>();
800   assert(dialect && "LLVM dialect must be registered");
801 
802   auto llvmModule = llvm::CloneModule(dialect->getLLVMModule());
803   if (!llvmModule)
804     return nullptr;
805 
806   llvm::LLVMContext &llvmContext = llvmModule->getContext();
807   llvm::IRBuilder<> builder(llvmContext);
808 
809   // Inject declarations for `malloc` and `free` functions that can be used in
810   // memref allocation/deallocation coming from standard ops lowering.
811   llvmModule->getOrInsertFunction("malloc", builder.getInt8PtrTy(),
812                                   builder.getInt64Ty());
813   llvmModule->getOrInsertFunction("free", builder.getVoidTy(),
814                                   builder.getInt8PtrTy());
815 
816   return llvmModule;
817 }
818