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