xref: /freebsd/contrib/llvm-project/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp (revision 7ef62cebc2f965b0f640263e179276928885e33d)
1 //===- OpenMPIRBuilder.cpp - Builder for LLVM-IR for OpenMP directives ----===//
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 /// \file
9 ///
10 /// This file implements the OpenMPIRBuilder class, which is used as a
11 /// convenient way to create LLVM instructions for OpenMP directives.
12 ///
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/Frontend/OpenMP/OMPIRBuilder.h"
16 #include "llvm/ADT/SmallSet.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/CodeMetrics.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
22 #include "llvm/Analysis/ScalarEvolution.h"
23 #include "llvm/Analysis/TargetLibraryInfo.h"
24 #include "llvm/IR/CFG.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DebugInfoMetadata.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/MDBuilder.h"
31 #include "llvm/IR/PassManager.h"
32 #include "llvm/IR/Value.h"
33 #include "llvm/MC/TargetRegistry.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Target/TargetMachine.h"
36 #include "llvm/Target/TargetOptions.h"
37 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
38 #include "llvm/Transforms/Utils/Cloning.h"
39 #include "llvm/Transforms/Utils/CodeExtractor.h"
40 #include "llvm/Transforms/Utils/LoopPeel.h"
41 #include "llvm/Transforms/Utils/UnrollLoop.h"
42 
43 #include <cstdint>
44 #include <optional>
45 
46 #define DEBUG_TYPE "openmp-ir-builder"
47 
48 using namespace llvm;
49 using namespace omp;
50 
51 static cl::opt<bool>
52     OptimisticAttributes("openmp-ir-builder-optimistic-attributes", cl::Hidden,
53                          cl::desc("Use optimistic attributes describing "
54                                   "'as-if' properties of runtime calls."),
55                          cl::init(false));
56 
57 static cl::opt<double> UnrollThresholdFactor(
58     "openmp-ir-builder-unroll-threshold-factor", cl::Hidden,
59     cl::desc("Factor for the unroll threshold to account for code "
60              "simplifications still taking place"),
61     cl::init(1.5));
62 
63 #ifndef NDEBUG
64 /// Return whether IP1 and IP2 are ambiguous, i.e. that inserting instructions
65 /// at position IP1 may change the meaning of IP2 or vice-versa. This is because
66 /// an InsertPoint stores the instruction before something is inserted. For
67 /// instance, if both point to the same instruction, two IRBuilders alternating
68 /// creating instruction will cause the instructions to be interleaved.
69 static bool isConflictIP(IRBuilder<>::InsertPoint IP1,
70                          IRBuilder<>::InsertPoint IP2) {
71   if (!IP1.isSet() || !IP2.isSet())
72     return false;
73   return IP1.getBlock() == IP2.getBlock() && IP1.getPoint() == IP2.getPoint();
74 }
75 
76 static bool isValidWorkshareLoopScheduleType(OMPScheduleType SchedType) {
77   // Valid ordered/unordered and base algorithm combinations.
78   switch (SchedType & ~OMPScheduleType::MonotonicityMask) {
79   case OMPScheduleType::UnorderedStaticChunked:
80   case OMPScheduleType::UnorderedStatic:
81   case OMPScheduleType::UnorderedDynamicChunked:
82   case OMPScheduleType::UnorderedGuidedChunked:
83   case OMPScheduleType::UnorderedRuntime:
84   case OMPScheduleType::UnorderedAuto:
85   case OMPScheduleType::UnorderedTrapezoidal:
86   case OMPScheduleType::UnorderedGreedy:
87   case OMPScheduleType::UnorderedBalanced:
88   case OMPScheduleType::UnorderedGuidedIterativeChunked:
89   case OMPScheduleType::UnorderedGuidedAnalyticalChunked:
90   case OMPScheduleType::UnorderedSteal:
91   case OMPScheduleType::UnorderedStaticBalancedChunked:
92   case OMPScheduleType::UnorderedGuidedSimd:
93   case OMPScheduleType::UnorderedRuntimeSimd:
94   case OMPScheduleType::OrderedStaticChunked:
95   case OMPScheduleType::OrderedStatic:
96   case OMPScheduleType::OrderedDynamicChunked:
97   case OMPScheduleType::OrderedGuidedChunked:
98   case OMPScheduleType::OrderedRuntime:
99   case OMPScheduleType::OrderedAuto:
100   case OMPScheduleType::OrderdTrapezoidal:
101   case OMPScheduleType::NomergeUnorderedStaticChunked:
102   case OMPScheduleType::NomergeUnorderedStatic:
103   case OMPScheduleType::NomergeUnorderedDynamicChunked:
104   case OMPScheduleType::NomergeUnorderedGuidedChunked:
105   case OMPScheduleType::NomergeUnorderedRuntime:
106   case OMPScheduleType::NomergeUnorderedAuto:
107   case OMPScheduleType::NomergeUnorderedTrapezoidal:
108   case OMPScheduleType::NomergeUnorderedGreedy:
109   case OMPScheduleType::NomergeUnorderedBalanced:
110   case OMPScheduleType::NomergeUnorderedGuidedIterativeChunked:
111   case OMPScheduleType::NomergeUnorderedGuidedAnalyticalChunked:
112   case OMPScheduleType::NomergeUnorderedSteal:
113   case OMPScheduleType::NomergeOrderedStaticChunked:
114   case OMPScheduleType::NomergeOrderedStatic:
115   case OMPScheduleType::NomergeOrderedDynamicChunked:
116   case OMPScheduleType::NomergeOrderedGuidedChunked:
117   case OMPScheduleType::NomergeOrderedRuntime:
118   case OMPScheduleType::NomergeOrderedAuto:
119   case OMPScheduleType::NomergeOrderedTrapezoidal:
120     break;
121   default:
122     return false;
123   }
124 
125   // Must not set both monotonicity modifiers at the same time.
126   OMPScheduleType MonotonicityFlags =
127       SchedType & OMPScheduleType::MonotonicityMask;
128   if (MonotonicityFlags == OMPScheduleType::MonotonicityMask)
129     return false;
130 
131   return true;
132 }
133 #endif
134 
135 /// Determine which scheduling algorithm to use, determined from schedule clause
136 /// arguments.
137 static OMPScheduleType
138 getOpenMPBaseScheduleType(llvm::omp::ScheduleKind ClauseKind, bool HasChunks,
139                           bool HasSimdModifier) {
140   // Currently, the default schedule it static.
141   switch (ClauseKind) {
142   case OMP_SCHEDULE_Default:
143   case OMP_SCHEDULE_Static:
144     return HasChunks ? OMPScheduleType::BaseStaticChunked
145                      : OMPScheduleType::BaseStatic;
146   case OMP_SCHEDULE_Dynamic:
147     return OMPScheduleType::BaseDynamicChunked;
148   case OMP_SCHEDULE_Guided:
149     return HasSimdModifier ? OMPScheduleType::BaseGuidedSimd
150                            : OMPScheduleType::BaseGuidedChunked;
151   case OMP_SCHEDULE_Auto:
152     return llvm::omp::OMPScheduleType::BaseAuto;
153   case OMP_SCHEDULE_Runtime:
154     return HasSimdModifier ? OMPScheduleType::BaseRuntimeSimd
155                            : OMPScheduleType::BaseRuntime;
156   }
157   llvm_unreachable("unhandled schedule clause argument");
158 }
159 
160 /// Adds ordering modifier flags to schedule type.
161 static OMPScheduleType
162 getOpenMPOrderingScheduleType(OMPScheduleType BaseScheduleType,
163                               bool HasOrderedClause) {
164   assert((BaseScheduleType & OMPScheduleType::ModifierMask) ==
165              OMPScheduleType::None &&
166          "Must not have ordering nor monotonicity flags already set");
167 
168   OMPScheduleType OrderingModifier = HasOrderedClause
169                                          ? OMPScheduleType::ModifierOrdered
170                                          : OMPScheduleType::ModifierUnordered;
171   OMPScheduleType OrderingScheduleType = BaseScheduleType | OrderingModifier;
172 
173   // Unsupported combinations
174   if (OrderingScheduleType ==
175       (OMPScheduleType::BaseGuidedSimd | OMPScheduleType::ModifierOrdered))
176     return OMPScheduleType::OrderedGuidedChunked;
177   else if (OrderingScheduleType == (OMPScheduleType::BaseRuntimeSimd |
178                                     OMPScheduleType::ModifierOrdered))
179     return OMPScheduleType::OrderedRuntime;
180 
181   return OrderingScheduleType;
182 }
183 
184 /// Adds monotonicity modifier flags to schedule type.
185 static OMPScheduleType
186 getOpenMPMonotonicityScheduleType(OMPScheduleType ScheduleType,
187                                   bool HasSimdModifier, bool HasMonotonic,
188                                   bool HasNonmonotonic, bool HasOrderedClause) {
189   assert((ScheduleType & OMPScheduleType::MonotonicityMask) ==
190              OMPScheduleType::None &&
191          "Must not have monotonicity flags already set");
192   assert((!HasMonotonic || !HasNonmonotonic) &&
193          "Monotonic and Nonmonotonic are contradicting each other");
194 
195   if (HasMonotonic) {
196     return ScheduleType | OMPScheduleType::ModifierMonotonic;
197   } else if (HasNonmonotonic) {
198     return ScheduleType | OMPScheduleType::ModifierNonmonotonic;
199   } else {
200     // OpenMP 5.1, 2.11.4 Worksharing-Loop Construct, Description.
201     // If the static schedule kind is specified or if the ordered clause is
202     // specified, and if the nonmonotonic modifier is not specified, the
203     // effect is as if the monotonic modifier is specified. Otherwise, unless
204     // the monotonic modifier is specified, the effect is as if the
205     // nonmonotonic modifier is specified.
206     OMPScheduleType BaseScheduleType =
207         ScheduleType & ~OMPScheduleType::ModifierMask;
208     if ((BaseScheduleType == OMPScheduleType::BaseStatic) ||
209         (BaseScheduleType == OMPScheduleType::BaseStaticChunked) ||
210         HasOrderedClause) {
211       // The monotonic is used by default in openmp runtime library, so no need
212       // to set it.
213       return ScheduleType;
214     } else {
215       return ScheduleType | OMPScheduleType::ModifierNonmonotonic;
216     }
217   }
218 }
219 
220 /// Determine the schedule type using schedule and ordering clause arguments.
221 static OMPScheduleType
222 computeOpenMPScheduleType(ScheduleKind ClauseKind, bool HasChunks,
223                           bool HasSimdModifier, bool HasMonotonicModifier,
224                           bool HasNonmonotonicModifier, bool HasOrderedClause) {
225   OMPScheduleType BaseSchedule =
226       getOpenMPBaseScheduleType(ClauseKind, HasChunks, HasSimdModifier);
227   OMPScheduleType OrderedSchedule =
228       getOpenMPOrderingScheduleType(BaseSchedule, HasOrderedClause);
229   OMPScheduleType Result = getOpenMPMonotonicityScheduleType(
230       OrderedSchedule, HasSimdModifier, HasMonotonicModifier,
231       HasNonmonotonicModifier, HasOrderedClause);
232 
233   assert(isValidWorkshareLoopScheduleType(Result));
234   return Result;
235 }
236 
237 /// Make \p Source branch to \p Target.
238 ///
239 /// Handles two situations:
240 /// * \p Source already has an unconditional branch.
241 /// * \p Source is a degenerate block (no terminator because the BB is
242 ///             the current head of the IR construction).
243 static void redirectTo(BasicBlock *Source, BasicBlock *Target, DebugLoc DL) {
244   if (Instruction *Term = Source->getTerminator()) {
245     auto *Br = cast<BranchInst>(Term);
246     assert(!Br->isConditional() &&
247            "BB's terminator must be an unconditional branch (or degenerate)");
248     BasicBlock *Succ = Br->getSuccessor(0);
249     Succ->removePredecessor(Source, /*KeepOneInputPHIs=*/true);
250     Br->setSuccessor(0, Target);
251     return;
252   }
253 
254   auto *NewBr = BranchInst::Create(Target, Source);
255   NewBr->setDebugLoc(DL);
256 }
257 
258 void llvm::spliceBB(IRBuilderBase::InsertPoint IP, BasicBlock *New,
259                     bool CreateBranch) {
260   assert(New->getFirstInsertionPt() == New->begin() &&
261          "Target BB must not have PHI nodes");
262 
263   // Move instructions to new block.
264   BasicBlock *Old = IP.getBlock();
265   New->splice(New->begin(), Old, IP.getPoint(), Old->end());
266 
267   if (CreateBranch)
268     BranchInst::Create(New, Old);
269 }
270 
271 void llvm::spliceBB(IRBuilder<> &Builder, BasicBlock *New, bool CreateBranch) {
272   DebugLoc DebugLoc = Builder.getCurrentDebugLocation();
273   BasicBlock *Old = Builder.GetInsertBlock();
274 
275   spliceBB(Builder.saveIP(), New, CreateBranch);
276   if (CreateBranch)
277     Builder.SetInsertPoint(Old->getTerminator());
278   else
279     Builder.SetInsertPoint(Old);
280 
281   // SetInsertPoint also updates the Builder's debug location, but we want to
282   // keep the one the Builder was configured to use.
283   Builder.SetCurrentDebugLocation(DebugLoc);
284 }
285 
286 BasicBlock *llvm::splitBB(IRBuilderBase::InsertPoint IP, bool CreateBranch,
287                           llvm::Twine Name) {
288   BasicBlock *Old = IP.getBlock();
289   BasicBlock *New = BasicBlock::Create(
290       Old->getContext(), Name.isTriviallyEmpty() ? Old->getName() : Name,
291       Old->getParent(), Old->getNextNode());
292   spliceBB(IP, New, CreateBranch);
293   New->replaceSuccessorsPhiUsesWith(Old, New);
294   return New;
295 }
296 
297 BasicBlock *llvm::splitBB(IRBuilderBase &Builder, bool CreateBranch,
298                           llvm::Twine Name) {
299   DebugLoc DebugLoc = Builder.getCurrentDebugLocation();
300   BasicBlock *New = splitBB(Builder.saveIP(), CreateBranch, Name);
301   if (CreateBranch)
302     Builder.SetInsertPoint(Builder.GetInsertBlock()->getTerminator());
303   else
304     Builder.SetInsertPoint(Builder.GetInsertBlock());
305   // SetInsertPoint also updates the Builder's debug location, but we want to
306   // keep the one the Builder was configured to use.
307   Builder.SetCurrentDebugLocation(DebugLoc);
308   return New;
309 }
310 
311 BasicBlock *llvm::splitBB(IRBuilder<> &Builder, bool CreateBranch,
312                           llvm::Twine Name) {
313   DebugLoc DebugLoc = Builder.getCurrentDebugLocation();
314   BasicBlock *New = splitBB(Builder.saveIP(), CreateBranch, Name);
315   if (CreateBranch)
316     Builder.SetInsertPoint(Builder.GetInsertBlock()->getTerminator());
317   else
318     Builder.SetInsertPoint(Builder.GetInsertBlock());
319   // SetInsertPoint also updates the Builder's debug location, but we want to
320   // keep the one the Builder was configured to use.
321   Builder.SetCurrentDebugLocation(DebugLoc);
322   return New;
323 }
324 
325 BasicBlock *llvm::splitBBWithSuffix(IRBuilderBase &Builder, bool CreateBranch,
326                                     llvm::Twine Suffix) {
327   BasicBlock *Old = Builder.GetInsertBlock();
328   return splitBB(Builder, CreateBranch, Old->getName() + Suffix);
329 }
330 
331 void OpenMPIRBuilder::addAttributes(omp::RuntimeFunction FnID, Function &Fn) {
332   LLVMContext &Ctx = Fn.getContext();
333   Triple T(M.getTargetTriple());
334 
335   // Get the function's current attributes.
336   auto Attrs = Fn.getAttributes();
337   auto FnAttrs = Attrs.getFnAttrs();
338   auto RetAttrs = Attrs.getRetAttrs();
339   SmallVector<AttributeSet, 4> ArgAttrs;
340   for (size_t ArgNo = 0; ArgNo < Fn.arg_size(); ++ArgNo)
341     ArgAttrs.emplace_back(Attrs.getParamAttrs(ArgNo));
342 
343   // Add AS to FnAS while taking special care with integer extensions.
344   auto addAttrSet = [&](AttributeSet &FnAS, const AttributeSet &AS,
345                         bool Param = true) -> void {
346     bool HasSignExt = AS.hasAttribute(Attribute::SExt);
347     bool HasZeroExt = AS.hasAttribute(Attribute::ZExt);
348     if (HasSignExt || HasZeroExt) {
349       assert(AS.getNumAttributes() == 1 &&
350              "Currently not handling extension attr combined with others.");
351       if (Param) {
352         if (auto AK = TargetLibraryInfo::getExtAttrForI32Param(T, HasSignExt))
353           FnAS = FnAS.addAttribute(Ctx, AK);
354       } else
355         if (auto AK = TargetLibraryInfo::getExtAttrForI32Return(T, HasSignExt))
356           FnAS = FnAS.addAttribute(Ctx, AK);
357     } else {
358       FnAS = FnAS.addAttributes(Ctx, AS);
359     }
360   };
361 
362 #define OMP_ATTRS_SET(VarName, AttrSet) AttributeSet VarName = AttrSet;
363 #include "llvm/Frontend/OpenMP/OMPKinds.def"
364 
365   // Add attributes to the function declaration.
366   switch (FnID) {
367 #define OMP_RTL_ATTRS(Enum, FnAttrSet, RetAttrSet, ArgAttrSets)                \
368   case Enum:                                                                   \
369     FnAttrs = FnAttrs.addAttributes(Ctx, FnAttrSet);                           \
370     addAttrSet(RetAttrs, RetAttrSet, /*Param*/false);                          \
371     for (size_t ArgNo = 0; ArgNo < ArgAttrSets.size(); ++ArgNo)                \
372       addAttrSet(ArgAttrs[ArgNo], ArgAttrSets[ArgNo]);                         \
373     Fn.setAttributes(AttributeList::get(Ctx, FnAttrs, RetAttrs, ArgAttrs));    \
374     break;
375 #include "llvm/Frontend/OpenMP/OMPKinds.def"
376   default:
377     // Attributes are optional.
378     break;
379   }
380 }
381 
382 FunctionCallee
383 OpenMPIRBuilder::getOrCreateRuntimeFunction(Module &M, RuntimeFunction FnID) {
384   FunctionType *FnTy = nullptr;
385   Function *Fn = nullptr;
386 
387   // Try to find the declation in the module first.
388   switch (FnID) {
389 #define OMP_RTL(Enum, Str, IsVarArg, ReturnType, ...)                          \
390   case Enum:                                                                   \
391     FnTy = FunctionType::get(ReturnType, ArrayRef<Type *>{__VA_ARGS__},        \
392                              IsVarArg);                                        \
393     Fn = M.getFunction(Str);                                                   \
394     break;
395 #include "llvm/Frontend/OpenMP/OMPKinds.def"
396   }
397 
398   if (!Fn) {
399     // Create a new declaration if we need one.
400     switch (FnID) {
401 #define OMP_RTL(Enum, Str, ...)                                                \
402   case Enum:                                                                   \
403     Fn = Function::Create(FnTy, GlobalValue::ExternalLinkage, Str, M);         \
404     break;
405 #include "llvm/Frontend/OpenMP/OMPKinds.def"
406     }
407 
408     // Add information if the runtime function takes a callback function
409     if (FnID == OMPRTL___kmpc_fork_call || FnID == OMPRTL___kmpc_fork_teams) {
410       if (!Fn->hasMetadata(LLVMContext::MD_callback)) {
411         LLVMContext &Ctx = Fn->getContext();
412         MDBuilder MDB(Ctx);
413         // Annotate the callback behavior of the runtime function:
414         //  - The callback callee is argument number 2 (microtask).
415         //  - The first two arguments of the callback callee are unknown (-1).
416         //  - All variadic arguments to the runtime function are passed to the
417         //    callback callee.
418         Fn->addMetadata(
419             LLVMContext::MD_callback,
420             *MDNode::get(Ctx, {MDB.createCallbackEncoding(
421                                   2, {-1, -1}, /* VarArgsArePassed */ true)}));
422       }
423     }
424 
425     LLVM_DEBUG(dbgs() << "Created OpenMP runtime function " << Fn->getName()
426                       << " with type " << *Fn->getFunctionType() << "\n");
427     addAttributes(FnID, *Fn);
428 
429   } else {
430     LLVM_DEBUG(dbgs() << "Found OpenMP runtime function " << Fn->getName()
431                       << " with type " << *Fn->getFunctionType() << "\n");
432   }
433 
434   assert(Fn && "Failed to create OpenMP runtime function");
435 
436   // Cast the function to the expected type if necessary
437   Constant *C = ConstantExpr::getBitCast(Fn, FnTy->getPointerTo());
438   return {FnTy, C};
439 }
440 
441 Function *OpenMPIRBuilder::getOrCreateRuntimeFunctionPtr(RuntimeFunction FnID) {
442   FunctionCallee RTLFn = getOrCreateRuntimeFunction(M, FnID);
443   auto *Fn = dyn_cast<llvm::Function>(RTLFn.getCallee());
444   assert(Fn && "Failed to create OpenMP runtime function pointer");
445   return Fn;
446 }
447 
448 void OpenMPIRBuilder::initialize() { initializeTypes(M); }
449 
450 void OpenMPIRBuilder::finalize(Function *Fn) {
451   SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet;
452   SmallVector<BasicBlock *, 32> Blocks;
453   SmallVector<OutlineInfo, 16> DeferredOutlines;
454   for (OutlineInfo &OI : OutlineInfos) {
455     // Skip functions that have not finalized yet; may happen with nested
456     // function generation.
457     if (Fn && OI.getFunction() != Fn) {
458       DeferredOutlines.push_back(OI);
459       continue;
460     }
461 
462     ParallelRegionBlockSet.clear();
463     Blocks.clear();
464     OI.collectBlocks(ParallelRegionBlockSet, Blocks);
465 
466     Function *OuterFn = OI.getFunction();
467     CodeExtractorAnalysisCache CEAC(*OuterFn);
468     CodeExtractor Extractor(Blocks, /* DominatorTree */ nullptr,
469                             /* AggregateArgs */ true,
470                             /* BlockFrequencyInfo */ nullptr,
471                             /* BranchProbabilityInfo */ nullptr,
472                             /* AssumptionCache */ nullptr,
473                             /* AllowVarArgs */ true,
474                             /* AllowAlloca */ true,
475                             /* AllocaBlock*/ OI.OuterAllocaBB,
476                             /* Suffix */ ".omp_par");
477 
478     LLVM_DEBUG(dbgs() << "Before     outlining: " << *OuterFn << "\n");
479     LLVM_DEBUG(dbgs() << "Entry " << OI.EntryBB->getName()
480                       << " Exit: " << OI.ExitBB->getName() << "\n");
481     assert(Extractor.isEligible() &&
482            "Expected OpenMP outlining to be possible!");
483 
484     for (auto *V : OI.ExcludeArgsFromAggregate)
485       Extractor.excludeArgFromAggregate(V);
486 
487     Function *OutlinedFn = Extractor.extractCodeRegion(CEAC);
488 
489     LLVM_DEBUG(dbgs() << "After      outlining: " << *OuterFn << "\n");
490     LLVM_DEBUG(dbgs() << "   Outlined function: " << *OutlinedFn << "\n");
491     assert(OutlinedFn->getReturnType()->isVoidTy() &&
492            "OpenMP outlined functions should not return a value!");
493 
494     // For compability with the clang CG we move the outlined function after the
495     // one with the parallel region.
496     OutlinedFn->removeFromParent();
497     M.getFunctionList().insertAfter(OuterFn->getIterator(), OutlinedFn);
498 
499     // Remove the artificial entry introduced by the extractor right away, we
500     // made our own entry block after all.
501     {
502       BasicBlock &ArtificialEntry = OutlinedFn->getEntryBlock();
503       assert(ArtificialEntry.getUniqueSuccessor() == OI.EntryBB);
504       assert(OI.EntryBB->getUniquePredecessor() == &ArtificialEntry);
505       // Move instructions from the to-be-deleted ArtificialEntry to the entry
506       // basic block of the parallel region. CodeExtractor generates
507       // instructions to unwrap the aggregate argument and may sink
508       // allocas/bitcasts for values that are solely used in the outlined region
509       // and do not escape.
510       assert(!ArtificialEntry.empty() &&
511              "Expected instructions to add in the outlined region entry");
512       for (BasicBlock::reverse_iterator It = ArtificialEntry.rbegin(),
513                                         End = ArtificialEntry.rend();
514            It != End;) {
515         Instruction &I = *It;
516         It++;
517 
518         if (I.isTerminator())
519           continue;
520 
521         I.moveBefore(*OI.EntryBB, OI.EntryBB->getFirstInsertionPt());
522       }
523 
524       OI.EntryBB->moveBefore(&ArtificialEntry);
525       ArtificialEntry.eraseFromParent();
526     }
527     assert(&OutlinedFn->getEntryBlock() == OI.EntryBB);
528     assert(OutlinedFn && OutlinedFn->getNumUses() == 1);
529 
530     // Run a user callback, e.g. to add attributes.
531     if (OI.PostOutlineCB)
532       OI.PostOutlineCB(*OutlinedFn);
533   }
534 
535   // Remove work items that have been completed.
536   OutlineInfos = std::move(DeferredOutlines);
537 }
538 
539 OpenMPIRBuilder::~OpenMPIRBuilder() {
540   assert(OutlineInfos.empty() && "There must be no outstanding outlinings");
541 }
542 
543 GlobalValue *OpenMPIRBuilder::createGlobalFlag(unsigned Value, StringRef Name) {
544   IntegerType *I32Ty = Type::getInt32Ty(M.getContext());
545   auto *GV =
546       new GlobalVariable(M, I32Ty,
547                          /* isConstant = */ true, GlobalValue::WeakODRLinkage,
548                          ConstantInt::get(I32Ty, Value), Name);
549   GV->setVisibility(GlobalValue::HiddenVisibility);
550 
551   return GV;
552 }
553 
554 Constant *OpenMPIRBuilder::getOrCreateIdent(Constant *SrcLocStr,
555                                             uint32_t SrcLocStrSize,
556                                             IdentFlag LocFlags,
557                                             unsigned Reserve2Flags) {
558   // Enable "C-mode".
559   LocFlags |= OMP_IDENT_FLAG_KMPC;
560 
561   Constant *&Ident =
562       IdentMap[{SrcLocStr, uint64_t(LocFlags) << 31 | Reserve2Flags}];
563   if (!Ident) {
564     Constant *I32Null = ConstantInt::getNullValue(Int32);
565     Constant *IdentData[] = {I32Null,
566                              ConstantInt::get(Int32, uint32_t(LocFlags)),
567                              ConstantInt::get(Int32, Reserve2Flags),
568                              ConstantInt::get(Int32, SrcLocStrSize), SrcLocStr};
569     Constant *Initializer =
570         ConstantStruct::get(OpenMPIRBuilder::Ident, IdentData);
571 
572     // Look for existing encoding of the location + flags, not needed but
573     // minimizes the difference to the existing solution while we transition.
574     for (GlobalVariable &GV : M.getGlobalList())
575       if (GV.getValueType() == OpenMPIRBuilder::Ident && GV.hasInitializer())
576         if (GV.getInitializer() == Initializer)
577           Ident = &GV;
578 
579     if (!Ident) {
580       auto *GV = new GlobalVariable(
581           M, OpenMPIRBuilder::Ident,
582           /* isConstant = */ true, GlobalValue::PrivateLinkage, Initializer, "",
583           nullptr, GlobalValue::NotThreadLocal,
584           M.getDataLayout().getDefaultGlobalsAddressSpace());
585       GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
586       GV->setAlignment(Align(8));
587       Ident = GV;
588     }
589   }
590 
591   return ConstantExpr::getPointerBitCastOrAddrSpaceCast(Ident, IdentPtr);
592 }
593 
594 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef LocStr,
595                                                 uint32_t &SrcLocStrSize) {
596   SrcLocStrSize = LocStr.size();
597   Constant *&SrcLocStr = SrcLocStrMap[LocStr];
598   if (!SrcLocStr) {
599     Constant *Initializer =
600         ConstantDataArray::getString(M.getContext(), LocStr);
601 
602     // Look for existing encoding of the location, not needed but minimizes the
603     // difference to the existing solution while we transition.
604     for (GlobalVariable &GV : M.getGlobalList())
605       if (GV.isConstant() && GV.hasInitializer() &&
606           GV.getInitializer() == Initializer)
607         return SrcLocStr = ConstantExpr::getPointerCast(&GV, Int8Ptr);
608 
609     SrcLocStr = Builder.CreateGlobalStringPtr(LocStr, /* Name */ "",
610                                               /* AddressSpace */ 0, &M);
611   }
612   return SrcLocStr;
613 }
614 
615 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef FunctionName,
616                                                 StringRef FileName,
617                                                 unsigned Line, unsigned Column,
618                                                 uint32_t &SrcLocStrSize) {
619   SmallString<128> Buffer;
620   Buffer.push_back(';');
621   Buffer.append(FileName);
622   Buffer.push_back(';');
623   Buffer.append(FunctionName);
624   Buffer.push_back(';');
625   Buffer.append(std::to_string(Line));
626   Buffer.push_back(';');
627   Buffer.append(std::to_string(Column));
628   Buffer.push_back(';');
629   Buffer.push_back(';');
630   return getOrCreateSrcLocStr(Buffer.str(), SrcLocStrSize);
631 }
632 
633 Constant *
634 OpenMPIRBuilder::getOrCreateDefaultSrcLocStr(uint32_t &SrcLocStrSize) {
635   StringRef UnknownLoc = ";unknown;unknown;0;0;;";
636   return getOrCreateSrcLocStr(UnknownLoc, SrcLocStrSize);
637 }
638 
639 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(DebugLoc DL,
640                                                 uint32_t &SrcLocStrSize,
641                                                 Function *F) {
642   DILocation *DIL = DL.get();
643   if (!DIL)
644     return getOrCreateDefaultSrcLocStr(SrcLocStrSize);
645   StringRef FileName = M.getName();
646   if (DIFile *DIF = DIL->getFile())
647     if (std::optional<StringRef> Source = DIF->getSource())
648       FileName = *Source;
649   StringRef Function = DIL->getScope()->getSubprogram()->getName();
650   if (Function.empty() && F)
651     Function = F->getName();
652   return getOrCreateSrcLocStr(Function, FileName, DIL->getLine(),
653                               DIL->getColumn(), SrcLocStrSize);
654 }
655 
656 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(const LocationDescription &Loc,
657                                                 uint32_t &SrcLocStrSize) {
658   return getOrCreateSrcLocStr(Loc.DL, SrcLocStrSize,
659                               Loc.IP.getBlock()->getParent());
660 }
661 
662 Value *OpenMPIRBuilder::getOrCreateThreadID(Value *Ident) {
663   return Builder.CreateCall(
664       getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_global_thread_num), Ident,
665       "omp_global_thread_num");
666 }
667 
668 OpenMPIRBuilder::InsertPointTy
669 OpenMPIRBuilder::createBarrier(const LocationDescription &Loc, Directive DK,
670                                bool ForceSimpleCall, bool CheckCancelFlag) {
671   if (!updateToLocation(Loc))
672     return Loc.IP;
673   return emitBarrierImpl(Loc, DK, ForceSimpleCall, CheckCancelFlag);
674 }
675 
676 OpenMPIRBuilder::InsertPointTy
677 OpenMPIRBuilder::emitBarrierImpl(const LocationDescription &Loc, Directive Kind,
678                                  bool ForceSimpleCall, bool CheckCancelFlag) {
679   // Build call __kmpc_cancel_barrier(loc, thread_id) or
680   //            __kmpc_barrier(loc, thread_id);
681 
682   IdentFlag BarrierLocFlags;
683   switch (Kind) {
684   case OMPD_for:
685     BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_FOR;
686     break;
687   case OMPD_sections:
688     BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SECTIONS;
689     break;
690   case OMPD_single:
691     BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SINGLE;
692     break;
693   case OMPD_barrier:
694     BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_EXPL;
695     break;
696   default:
697     BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL;
698     break;
699   }
700 
701   uint32_t SrcLocStrSize;
702   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
703   Value *Args[] = {
704       getOrCreateIdent(SrcLocStr, SrcLocStrSize, BarrierLocFlags),
705       getOrCreateThreadID(getOrCreateIdent(SrcLocStr, SrcLocStrSize))};
706 
707   // If we are in a cancellable parallel region, barriers are cancellation
708   // points.
709   // TODO: Check why we would force simple calls or to ignore the cancel flag.
710   bool UseCancelBarrier =
711       !ForceSimpleCall && isLastFinalizationInfoCancellable(OMPD_parallel);
712 
713   Value *Result =
714       Builder.CreateCall(getOrCreateRuntimeFunctionPtr(
715                              UseCancelBarrier ? OMPRTL___kmpc_cancel_barrier
716                                               : OMPRTL___kmpc_barrier),
717                          Args);
718 
719   if (UseCancelBarrier && CheckCancelFlag)
720     emitCancelationCheckImpl(Result, OMPD_parallel);
721 
722   return Builder.saveIP();
723 }
724 
725 OpenMPIRBuilder::InsertPointTy
726 OpenMPIRBuilder::createCancel(const LocationDescription &Loc,
727                               Value *IfCondition,
728                               omp::Directive CanceledDirective) {
729   if (!updateToLocation(Loc))
730     return Loc.IP;
731 
732   // LLVM utilities like blocks with terminators.
733   auto *UI = Builder.CreateUnreachable();
734 
735   Instruction *ThenTI = UI, *ElseTI = nullptr;
736   if (IfCondition)
737     SplitBlockAndInsertIfThenElse(IfCondition, UI, &ThenTI, &ElseTI);
738   Builder.SetInsertPoint(ThenTI);
739 
740   Value *CancelKind = nullptr;
741   switch (CanceledDirective) {
742 #define OMP_CANCEL_KIND(Enum, Str, DirectiveEnum, Value)                       \
743   case DirectiveEnum:                                                          \
744     CancelKind = Builder.getInt32(Value);                                      \
745     break;
746 #include "llvm/Frontend/OpenMP/OMPKinds.def"
747   default:
748     llvm_unreachable("Unknown cancel kind!");
749   }
750 
751   uint32_t SrcLocStrSize;
752   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
753   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
754   Value *Args[] = {Ident, getOrCreateThreadID(Ident), CancelKind};
755   Value *Result = Builder.CreateCall(
756       getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_cancel), Args);
757   auto ExitCB = [this, CanceledDirective, Loc](InsertPointTy IP) {
758     if (CanceledDirective == OMPD_parallel) {
759       IRBuilder<>::InsertPointGuard IPG(Builder);
760       Builder.restoreIP(IP);
761       createBarrier(LocationDescription(Builder.saveIP(), Loc.DL),
762                     omp::Directive::OMPD_unknown, /* ForceSimpleCall */ false,
763                     /* CheckCancelFlag */ false);
764     }
765   };
766 
767   // The actual cancel logic is shared with others, e.g., cancel_barriers.
768   emitCancelationCheckImpl(Result, CanceledDirective, ExitCB);
769 
770   // Update the insertion point and remove the terminator we introduced.
771   Builder.SetInsertPoint(UI->getParent());
772   UI->eraseFromParent();
773 
774   return Builder.saveIP();
775 }
776 
777 void OpenMPIRBuilder::emitOffloadingEntry(Constant *Addr, StringRef Name,
778                                           uint64_t Size, int32_t Flags,
779                                           StringRef SectionName) {
780   Type *Int8PtrTy = Type::getInt8PtrTy(M.getContext());
781   Type *Int32Ty = Type::getInt32Ty(M.getContext());
782   Type *SizeTy = M.getDataLayout().getIntPtrType(M.getContext());
783 
784   Constant *AddrName = ConstantDataArray::getString(M.getContext(), Name);
785 
786   // Create the constant string used to look up the symbol in the device.
787   auto *Str =
788       new llvm::GlobalVariable(M, AddrName->getType(), /*isConstant=*/true,
789                                llvm::GlobalValue::InternalLinkage, AddrName,
790                                ".omp_offloading.entry_name");
791   Str->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
792 
793   // Construct the offloading entry.
794   Constant *EntryData[] = {
795       ConstantExpr::getPointerBitCastOrAddrSpaceCast(Addr, Int8PtrTy),
796       ConstantExpr::getPointerBitCastOrAddrSpaceCast(Str, Int8PtrTy),
797       ConstantInt::get(SizeTy, Size),
798       ConstantInt::get(Int32Ty, Flags),
799       ConstantInt::get(Int32Ty, 0),
800   };
801   Constant *EntryInitializer =
802       ConstantStruct::get(OpenMPIRBuilder::OffloadEntry, EntryData);
803 
804   auto *Entry = new GlobalVariable(
805       M, OpenMPIRBuilder::OffloadEntry,
806       /* isConstant = */ true, GlobalValue::WeakAnyLinkage, EntryInitializer,
807       ".omp_offloading.entry." + Name, nullptr, GlobalValue::NotThreadLocal,
808       M.getDataLayout().getDefaultGlobalsAddressSpace());
809 
810   // The entry has to be created in the section the linker expects it to be.
811   Entry->setSection(SectionName);
812   Entry->setAlignment(Align(1));
813 }
814 
815 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitTargetKernel(
816     const LocationDescription &Loc, Value *&Return, Value *Ident,
817     Value *DeviceID, Value *NumTeams, Value *NumThreads, Value *HostPtr,
818     ArrayRef<Value *> KernelArgs) {
819   if (!updateToLocation(Loc))
820     return Loc.IP;
821 
822   auto *KernelArgsPtr =
823       Builder.CreateAlloca(OpenMPIRBuilder::KernelArgs, nullptr, "kernel_args");
824   for (unsigned I = 0, Size = KernelArgs.size(); I != Size; ++I) {
825     llvm::Value *Arg =
826         Builder.CreateStructGEP(OpenMPIRBuilder::KernelArgs, KernelArgsPtr, I);
827     Builder.CreateAlignedStore(
828         KernelArgs[I], Arg,
829         M.getDataLayout().getPrefTypeAlign(KernelArgs[I]->getType()));
830   }
831 
832   SmallVector<Value *> OffloadingArgs{Ident,      DeviceID, NumTeams,
833                                       NumThreads, HostPtr,  KernelArgsPtr};
834 
835   Return = Builder.CreateCall(
836       getOrCreateRuntimeFunction(M, OMPRTL___tgt_target_kernel),
837       OffloadingArgs);
838 
839   return Builder.saveIP();
840 }
841 
842 void OpenMPIRBuilder::emitCancelationCheckImpl(Value *CancelFlag,
843                                                omp::Directive CanceledDirective,
844                                                FinalizeCallbackTy ExitCB) {
845   assert(isLastFinalizationInfoCancellable(CanceledDirective) &&
846          "Unexpected cancellation!");
847 
848   // For a cancel barrier we create two new blocks.
849   BasicBlock *BB = Builder.GetInsertBlock();
850   BasicBlock *NonCancellationBlock;
851   if (Builder.GetInsertPoint() == BB->end()) {
852     // TODO: This branch will not be needed once we moved to the
853     // OpenMPIRBuilder codegen completely.
854     NonCancellationBlock = BasicBlock::Create(
855         BB->getContext(), BB->getName() + ".cont", BB->getParent());
856   } else {
857     NonCancellationBlock = SplitBlock(BB, &*Builder.GetInsertPoint());
858     BB->getTerminator()->eraseFromParent();
859     Builder.SetInsertPoint(BB);
860   }
861   BasicBlock *CancellationBlock = BasicBlock::Create(
862       BB->getContext(), BB->getName() + ".cncl", BB->getParent());
863 
864   // Jump to them based on the return value.
865   Value *Cmp = Builder.CreateIsNull(CancelFlag);
866   Builder.CreateCondBr(Cmp, NonCancellationBlock, CancellationBlock,
867                        /* TODO weight */ nullptr, nullptr);
868 
869   // From the cancellation block we finalize all variables and go to the
870   // post finalization block that is known to the FiniCB callback.
871   Builder.SetInsertPoint(CancellationBlock);
872   if (ExitCB)
873     ExitCB(Builder.saveIP());
874   auto &FI = FinalizationStack.back();
875   FI.FiniCB(Builder.saveIP());
876 
877   // The continuation block is where code generation continues.
878   Builder.SetInsertPoint(NonCancellationBlock, NonCancellationBlock->begin());
879 }
880 
881 IRBuilder<>::InsertPoint OpenMPIRBuilder::createParallel(
882     const LocationDescription &Loc, InsertPointTy OuterAllocaIP,
883     BodyGenCallbackTy BodyGenCB, PrivatizeCallbackTy PrivCB,
884     FinalizeCallbackTy FiniCB, Value *IfCondition, Value *NumThreads,
885     omp::ProcBindKind ProcBind, bool IsCancellable) {
886   assert(!isConflictIP(Loc.IP, OuterAllocaIP) && "IPs must not be ambiguous");
887 
888   if (!updateToLocation(Loc))
889     return Loc.IP;
890 
891   uint32_t SrcLocStrSize;
892   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
893   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
894   Value *ThreadID = getOrCreateThreadID(Ident);
895 
896   if (NumThreads) {
897     // Build call __kmpc_push_num_threads(&Ident, global_tid, num_threads)
898     Value *Args[] = {
899         Ident, ThreadID,
900         Builder.CreateIntCast(NumThreads, Int32, /*isSigned*/ false)};
901     Builder.CreateCall(
902         getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_push_num_threads), Args);
903   }
904 
905   if (ProcBind != OMP_PROC_BIND_default) {
906     // Build call __kmpc_push_proc_bind(&Ident, global_tid, proc_bind)
907     Value *Args[] = {
908         Ident, ThreadID,
909         ConstantInt::get(Int32, unsigned(ProcBind), /*isSigned=*/true)};
910     Builder.CreateCall(
911         getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_push_proc_bind), Args);
912   }
913 
914   BasicBlock *InsertBB = Builder.GetInsertBlock();
915   Function *OuterFn = InsertBB->getParent();
916 
917   // Save the outer alloca block because the insertion iterator may get
918   // invalidated and we still need this later.
919   BasicBlock *OuterAllocaBlock = OuterAllocaIP.getBlock();
920 
921   // Vector to remember instructions we used only during the modeling but which
922   // we want to delete at the end.
923   SmallVector<Instruction *, 4> ToBeDeleted;
924 
925   // Change the location to the outer alloca insertion point to create and
926   // initialize the allocas we pass into the parallel region.
927   Builder.restoreIP(OuterAllocaIP);
928   AllocaInst *TIDAddr = Builder.CreateAlloca(Int32, nullptr, "tid.addr");
929   AllocaInst *ZeroAddr = Builder.CreateAlloca(Int32, nullptr, "zero.addr");
930 
931   // We only need TIDAddr and ZeroAddr for modeling purposes to get the
932   // associated arguments in the outlined function, so we delete them later.
933   ToBeDeleted.push_back(TIDAddr);
934   ToBeDeleted.push_back(ZeroAddr);
935 
936   // Create an artificial insertion point that will also ensure the blocks we
937   // are about to split are not degenerated.
938   auto *UI = new UnreachableInst(Builder.getContext(), InsertBB);
939 
940   BasicBlock *EntryBB = UI->getParent();
941   BasicBlock *PRegEntryBB = EntryBB->splitBasicBlock(UI, "omp.par.entry");
942   BasicBlock *PRegBodyBB = PRegEntryBB->splitBasicBlock(UI, "omp.par.region");
943   BasicBlock *PRegPreFiniBB =
944       PRegBodyBB->splitBasicBlock(UI, "omp.par.pre_finalize");
945   BasicBlock *PRegExitBB = PRegPreFiniBB->splitBasicBlock(UI, "omp.par.exit");
946 
947   auto FiniCBWrapper = [&](InsertPointTy IP) {
948     // Hide "open-ended" blocks from the given FiniCB by setting the right jump
949     // target to the region exit block.
950     if (IP.getBlock()->end() == IP.getPoint()) {
951       IRBuilder<>::InsertPointGuard IPG(Builder);
952       Builder.restoreIP(IP);
953       Instruction *I = Builder.CreateBr(PRegExitBB);
954       IP = InsertPointTy(I->getParent(), I->getIterator());
955     }
956     assert(IP.getBlock()->getTerminator()->getNumSuccessors() == 1 &&
957            IP.getBlock()->getTerminator()->getSuccessor(0) == PRegExitBB &&
958            "Unexpected insertion point for finalization call!");
959     return FiniCB(IP);
960   };
961 
962   FinalizationStack.push_back({FiniCBWrapper, OMPD_parallel, IsCancellable});
963 
964   // Generate the privatization allocas in the block that will become the entry
965   // of the outlined function.
966   Builder.SetInsertPoint(PRegEntryBB->getTerminator());
967   InsertPointTy InnerAllocaIP = Builder.saveIP();
968 
969   AllocaInst *PrivTIDAddr =
970       Builder.CreateAlloca(Int32, nullptr, "tid.addr.local");
971   Instruction *PrivTID = Builder.CreateLoad(Int32, PrivTIDAddr, "tid");
972 
973   // Add some fake uses for OpenMP provided arguments.
974   ToBeDeleted.push_back(Builder.CreateLoad(Int32, TIDAddr, "tid.addr.use"));
975   Instruction *ZeroAddrUse =
976       Builder.CreateLoad(Int32, ZeroAddr, "zero.addr.use");
977   ToBeDeleted.push_back(ZeroAddrUse);
978 
979   // EntryBB
980   //   |
981   //   V
982   // PRegionEntryBB         <- Privatization allocas are placed here.
983   //   |
984   //   V
985   // PRegionBodyBB          <- BodeGen is invoked here.
986   //   |
987   //   V
988   // PRegPreFiniBB          <- The block we will start finalization from.
989   //   |
990   //   V
991   // PRegionExitBB          <- A common exit to simplify block collection.
992   //
993 
994   LLVM_DEBUG(dbgs() << "Before body codegen: " << *OuterFn << "\n");
995 
996   // Let the caller create the body.
997   assert(BodyGenCB && "Expected body generation callback!");
998   InsertPointTy CodeGenIP(PRegBodyBB, PRegBodyBB->begin());
999   BodyGenCB(InnerAllocaIP, CodeGenIP);
1000 
1001   LLVM_DEBUG(dbgs() << "After  body codegen: " << *OuterFn << "\n");
1002   FunctionCallee RTLFn;
1003   if (IfCondition)
1004     RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_fork_call_if);
1005   else
1006     RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_fork_call);
1007 
1008   if (auto *F = dyn_cast<llvm::Function>(RTLFn.getCallee())) {
1009     if (!F->hasMetadata(llvm::LLVMContext::MD_callback)) {
1010       llvm::LLVMContext &Ctx = F->getContext();
1011       MDBuilder MDB(Ctx);
1012       // Annotate the callback behavior of the __kmpc_fork_call:
1013       //  - The callback callee is argument number 2 (microtask).
1014       //  - The first two arguments of the callback callee are unknown (-1).
1015       //  - All variadic arguments to the __kmpc_fork_call are passed to the
1016       //    callback callee.
1017       F->addMetadata(
1018           llvm::LLVMContext::MD_callback,
1019           *llvm::MDNode::get(
1020               Ctx, {MDB.createCallbackEncoding(2, {-1, -1},
1021                                                /* VarArgsArePassed */ true)}));
1022     }
1023   }
1024 
1025   OutlineInfo OI;
1026   OI.PostOutlineCB = [=](Function &OutlinedFn) {
1027     // Add some known attributes.
1028     OutlinedFn.addParamAttr(0, Attribute::NoAlias);
1029     OutlinedFn.addParamAttr(1, Attribute::NoAlias);
1030     OutlinedFn.addFnAttr(Attribute::NoUnwind);
1031     OutlinedFn.addFnAttr(Attribute::NoRecurse);
1032 
1033     assert(OutlinedFn.arg_size() >= 2 &&
1034            "Expected at least tid and bounded tid as arguments");
1035     unsigned NumCapturedVars =
1036         OutlinedFn.arg_size() - /* tid & bounded tid */ 2;
1037 
1038     CallInst *CI = cast<CallInst>(OutlinedFn.user_back());
1039     CI->getParent()->setName("omp_parallel");
1040     Builder.SetInsertPoint(CI);
1041 
1042     // Build call __kmpc_fork_call[_if](Ident, n, microtask, var1, .., varn);
1043     Value *ForkCallArgs[] = {
1044         Ident, Builder.getInt32(NumCapturedVars),
1045         Builder.CreateBitCast(&OutlinedFn, ParallelTaskPtr)};
1046 
1047     SmallVector<Value *, 16> RealArgs;
1048     RealArgs.append(std::begin(ForkCallArgs), std::end(ForkCallArgs));
1049     if (IfCondition) {
1050       Value *Cond = Builder.CreateSExtOrTrunc(IfCondition,
1051                                               Type::getInt32Ty(M.getContext()));
1052       RealArgs.push_back(Cond);
1053     }
1054     RealArgs.append(CI->arg_begin() + /* tid & bound tid */ 2, CI->arg_end());
1055 
1056     // __kmpc_fork_call_if always expects a void ptr as the last argument
1057     // If there are no arguments, pass a null pointer.
1058     auto PtrTy = Type::getInt8PtrTy(M.getContext());
1059     if (IfCondition && NumCapturedVars == 0) {
1060       llvm::Value *Void = ConstantPointerNull::get(PtrTy);
1061       RealArgs.push_back(Void);
1062     }
1063     if (IfCondition && RealArgs.back()->getType() != PtrTy)
1064       RealArgs.back() = Builder.CreateBitCast(RealArgs.back(), PtrTy);
1065 
1066     Builder.CreateCall(RTLFn, RealArgs);
1067 
1068     LLVM_DEBUG(dbgs() << "With fork_call placed: "
1069                       << *Builder.GetInsertBlock()->getParent() << "\n");
1070 
1071     InsertPointTy ExitIP(PRegExitBB, PRegExitBB->end());
1072 
1073     // Initialize the local TID stack location with the argument value.
1074     Builder.SetInsertPoint(PrivTID);
1075     Function::arg_iterator OutlinedAI = OutlinedFn.arg_begin();
1076     Builder.CreateStore(Builder.CreateLoad(Int32, OutlinedAI), PrivTIDAddr);
1077 
1078     CI->eraseFromParent();
1079 
1080     for (Instruction *I : ToBeDeleted)
1081       I->eraseFromParent();
1082   };
1083 
1084   // Adjust the finalization stack, verify the adjustment, and call the
1085   // finalize function a last time to finalize values between the pre-fini
1086   // block and the exit block if we left the parallel "the normal way".
1087   auto FiniInfo = FinalizationStack.pop_back_val();
1088   (void)FiniInfo;
1089   assert(FiniInfo.DK == OMPD_parallel &&
1090          "Unexpected finalization stack state!");
1091 
1092   Instruction *PRegPreFiniTI = PRegPreFiniBB->getTerminator();
1093 
1094   InsertPointTy PreFiniIP(PRegPreFiniBB, PRegPreFiniTI->getIterator());
1095   FiniCB(PreFiniIP);
1096 
1097   OI.OuterAllocaBB = OuterAllocaBlock;
1098   OI.EntryBB = PRegEntryBB;
1099   OI.ExitBB = PRegExitBB;
1100 
1101   SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet;
1102   SmallVector<BasicBlock *, 32> Blocks;
1103   OI.collectBlocks(ParallelRegionBlockSet, Blocks);
1104 
1105   // Ensure a single exit node for the outlined region by creating one.
1106   // We might have multiple incoming edges to the exit now due to finalizations,
1107   // e.g., cancel calls that cause the control flow to leave the region.
1108   BasicBlock *PRegOutlinedExitBB = PRegExitBB;
1109   PRegExitBB = SplitBlock(PRegExitBB, &*PRegExitBB->getFirstInsertionPt());
1110   PRegOutlinedExitBB->setName("omp.par.outlined.exit");
1111   Blocks.push_back(PRegOutlinedExitBB);
1112 
1113   CodeExtractorAnalysisCache CEAC(*OuterFn);
1114   CodeExtractor Extractor(Blocks, /* DominatorTree */ nullptr,
1115                           /* AggregateArgs */ false,
1116                           /* BlockFrequencyInfo */ nullptr,
1117                           /* BranchProbabilityInfo */ nullptr,
1118                           /* AssumptionCache */ nullptr,
1119                           /* AllowVarArgs */ true,
1120                           /* AllowAlloca */ true,
1121                           /* AllocationBlock */ OuterAllocaBlock,
1122                           /* Suffix */ ".omp_par");
1123 
1124   // Find inputs to, outputs from the code region.
1125   BasicBlock *CommonExit = nullptr;
1126   SetVector<Value *> Inputs, Outputs, SinkingCands, HoistingCands;
1127   Extractor.findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit);
1128   Extractor.findInputsOutputs(Inputs, Outputs, SinkingCands);
1129 
1130   LLVM_DEBUG(dbgs() << "Before privatization: " << *OuterFn << "\n");
1131 
1132   FunctionCallee TIDRTLFn =
1133       getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_global_thread_num);
1134 
1135   auto PrivHelper = [&](Value &V) {
1136     if (&V == TIDAddr || &V == ZeroAddr) {
1137       OI.ExcludeArgsFromAggregate.push_back(&V);
1138       return;
1139     }
1140 
1141     SetVector<Use *> Uses;
1142     for (Use &U : V.uses())
1143       if (auto *UserI = dyn_cast<Instruction>(U.getUser()))
1144         if (ParallelRegionBlockSet.count(UserI->getParent()))
1145           Uses.insert(&U);
1146 
1147     // __kmpc_fork_call expects extra arguments as pointers. If the input
1148     // already has a pointer type, everything is fine. Otherwise, store the
1149     // value onto stack and load it back inside the to-be-outlined region. This
1150     // will ensure only the pointer will be passed to the function.
1151     // FIXME: if there are more than 15 trailing arguments, they must be
1152     // additionally packed in a struct.
1153     Value *Inner = &V;
1154     if (!V.getType()->isPointerTy()) {
1155       IRBuilder<>::InsertPointGuard Guard(Builder);
1156       LLVM_DEBUG(llvm::dbgs() << "Forwarding input as pointer: " << V << "\n");
1157 
1158       Builder.restoreIP(OuterAllocaIP);
1159       Value *Ptr =
1160           Builder.CreateAlloca(V.getType(), nullptr, V.getName() + ".reloaded");
1161 
1162       // Store to stack at end of the block that currently branches to the entry
1163       // block of the to-be-outlined region.
1164       Builder.SetInsertPoint(InsertBB,
1165                              InsertBB->getTerminator()->getIterator());
1166       Builder.CreateStore(&V, Ptr);
1167 
1168       // Load back next to allocations in the to-be-outlined region.
1169       Builder.restoreIP(InnerAllocaIP);
1170       Inner = Builder.CreateLoad(V.getType(), Ptr);
1171     }
1172 
1173     Value *ReplacementValue = nullptr;
1174     CallInst *CI = dyn_cast<CallInst>(&V);
1175     if (CI && CI->getCalledFunction() == TIDRTLFn.getCallee()) {
1176       ReplacementValue = PrivTID;
1177     } else {
1178       Builder.restoreIP(
1179           PrivCB(InnerAllocaIP, Builder.saveIP(), V, *Inner, ReplacementValue));
1180       assert(ReplacementValue &&
1181              "Expected copy/create callback to set replacement value!");
1182       if (ReplacementValue == &V)
1183         return;
1184     }
1185 
1186     for (Use *UPtr : Uses)
1187       UPtr->set(ReplacementValue);
1188   };
1189 
1190   // Reset the inner alloca insertion as it will be used for loading the values
1191   // wrapped into pointers before passing them into the to-be-outlined region.
1192   // Configure it to insert immediately after the fake use of zero address so
1193   // that they are available in the generated body and so that the
1194   // OpenMP-related values (thread ID and zero address pointers) remain leading
1195   // in the argument list.
1196   InnerAllocaIP = IRBuilder<>::InsertPoint(
1197       ZeroAddrUse->getParent(), ZeroAddrUse->getNextNode()->getIterator());
1198 
1199   // Reset the outer alloca insertion point to the entry of the relevant block
1200   // in case it was invalidated.
1201   OuterAllocaIP = IRBuilder<>::InsertPoint(
1202       OuterAllocaBlock, OuterAllocaBlock->getFirstInsertionPt());
1203 
1204   for (Value *Input : Inputs) {
1205     LLVM_DEBUG(dbgs() << "Captured input: " << *Input << "\n");
1206     PrivHelper(*Input);
1207   }
1208   LLVM_DEBUG({
1209     for (Value *Output : Outputs)
1210       LLVM_DEBUG(dbgs() << "Captured output: " << *Output << "\n");
1211   });
1212   assert(Outputs.empty() &&
1213          "OpenMP outlining should not produce live-out values!");
1214 
1215   LLVM_DEBUG(dbgs() << "After  privatization: " << *OuterFn << "\n");
1216   LLVM_DEBUG({
1217     for (auto *BB : Blocks)
1218       dbgs() << " PBR: " << BB->getName() << "\n";
1219   });
1220 
1221   // Register the outlined info.
1222   addOutlineInfo(std::move(OI));
1223 
1224   InsertPointTy AfterIP(UI->getParent(), UI->getParent()->end());
1225   UI->eraseFromParent();
1226 
1227   return AfterIP;
1228 }
1229 
1230 void OpenMPIRBuilder::emitFlush(const LocationDescription &Loc) {
1231   // Build call void __kmpc_flush(ident_t *loc)
1232   uint32_t SrcLocStrSize;
1233   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
1234   Value *Args[] = {getOrCreateIdent(SrcLocStr, SrcLocStrSize)};
1235 
1236   Builder.CreateCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_flush), Args);
1237 }
1238 
1239 void OpenMPIRBuilder::createFlush(const LocationDescription &Loc) {
1240   if (!updateToLocation(Loc))
1241     return;
1242   emitFlush(Loc);
1243 }
1244 
1245 void OpenMPIRBuilder::emitTaskwaitImpl(const LocationDescription &Loc) {
1246   // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32
1247   // global_tid);
1248   uint32_t SrcLocStrSize;
1249   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
1250   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
1251   Value *Args[] = {Ident, getOrCreateThreadID(Ident)};
1252 
1253   // Ignore return result until untied tasks are supported.
1254   Builder.CreateCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_taskwait),
1255                      Args);
1256 }
1257 
1258 void OpenMPIRBuilder::createTaskwait(const LocationDescription &Loc) {
1259   if (!updateToLocation(Loc))
1260     return;
1261   emitTaskwaitImpl(Loc);
1262 }
1263 
1264 void OpenMPIRBuilder::emitTaskyieldImpl(const LocationDescription &Loc) {
1265   // Build call __kmpc_omp_taskyield(loc, thread_id, 0);
1266   uint32_t SrcLocStrSize;
1267   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
1268   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
1269   Constant *I32Null = ConstantInt::getNullValue(Int32);
1270   Value *Args[] = {Ident, getOrCreateThreadID(Ident), I32Null};
1271 
1272   Builder.CreateCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_taskyield),
1273                      Args);
1274 }
1275 
1276 void OpenMPIRBuilder::createTaskyield(const LocationDescription &Loc) {
1277   if (!updateToLocation(Loc))
1278     return;
1279   emitTaskyieldImpl(Loc);
1280 }
1281 
1282 OpenMPIRBuilder::InsertPointTy
1283 OpenMPIRBuilder::createTask(const LocationDescription &Loc,
1284                             InsertPointTy AllocaIP, BodyGenCallbackTy BodyGenCB,
1285                             bool Tied, Value *Final, Value *IfCondition,
1286                             SmallVector<DependData> Dependencies) {
1287   if (!updateToLocation(Loc))
1288     return InsertPointTy();
1289 
1290   uint32_t SrcLocStrSize;
1291   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
1292   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
1293   // The current basic block is split into four basic blocks. After outlining,
1294   // they will be mapped as follows:
1295   // ```
1296   // def current_fn() {
1297   //   current_basic_block:
1298   //     br label %task.exit
1299   //   task.exit:
1300   //     ; instructions after task
1301   // }
1302   // def outlined_fn() {
1303   //   task.alloca:
1304   //     br label %task.body
1305   //   task.body:
1306   //     ret void
1307   // }
1308   // ```
1309   BasicBlock *TaskExitBB = splitBB(Builder, /*CreateBranch=*/true, "task.exit");
1310   BasicBlock *TaskBodyBB = splitBB(Builder, /*CreateBranch=*/true, "task.body");
1311   BasicBlock *TaskAllocaBB =
1312       splitBB(Builder, /*CreateBranch=*/true, "task.alloca");
1313 
1314   OutlineInfo OI;
1315   OI.EntryBB = TaskAllocaBB;
1316   OI.OuterAllocaBB = AllocaIP.getBlock();
1317   OI.ExitBB = TaskExitBB;
1318   OI.PostOutlineCB = [this, Ident, Tied, Final, IfCondition,
1319                       Dependencies](Function &OutlinedFn) {
1320     // The input IR here looks like the following-
1321     // ```
1322     // func @current_fn() {
1323     //   outlined_fn(%args)
1324     // }
1325     // func @outlined_fn(%args) { ... }
1326     // ```
1327     //
1328     // This is changed to the following-
1329     //
1330     // ```
1331     // func @current_fn() {
1332     //   runtime_call(..., wrapper_fn, ...)
1333     // }
1334     // func @wrapper_fn(..., %args) {
1335     //   outlined_fn(%args)
1336     // }
1337     // func @outlined_fn(%args) { ... }
1338     // ```
1339 
1340     // The stale call instruction will be replaced with a new call instruction
1341     // for runtime call with a wrapper function.
1342     assert(OutlinedFn.getNumUses() == 1 &&
1343            "there must be a single user for the outlined function");
1344     CallInst *StaleCI = cast<CallInst>(OutlinedFn.user_back());
1345 
1346     // HasTaskData is true if any variables are captured in the outlined region,
1347     // false otherwise.
1348     bool HasTaskData = StaleCI->arg_size() > 0;
1349     Builder.SetInsertPoint(StaleCI);
1350 
1351     // Gather the arguments for emitting the runtime call for
1352     // @__kmpc_omp_task_alloc
1353     Function *TaskAllocFn =
1354         getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task_alloc);
1355 
1356     // Arguments - `loc_ref` (Ident) and `gtid` (ThreadID)
1357     // call.
1358     Value *ThreadID = getOrCreateThreadID(Ident);
1359 
1360     // Argument - `flags`
1361     // Task is tied iff (Flags & 1) == 1.
1362     // Task is untied iff (Flags & 1) == 0.
1363     // Task is final iff (Flags & 2) == 2.
1364     // Task is not final iff (Flags & 2) == 0.
1365     // TODO: Handle the other flags.
1366     Value *Flags = Builder.getInt32(Tied);
1367     if (Final) {
1368       Value *FinalFlag =
1369           Builder.CreateSelect(Final, Builder.getInt32(2), Builder.getInt32(0));
1370       Flags = Builder.CreateOr(FinalFlag, Flags);
1371     }
1372 
1373     // Argument - `sizeof_kmp_task_t` (TaskSize)
1374     // Tasksize refers to the size in bytes of kmp_task_t data structure
1375     // including private vars accessed in task.
1376     Value *TaskSize = Builder.getInt64(0);
1377     if (HasTaskData) {
1378       AllocaInst *ArgStructAlloca =
1379           dyn_cast<AllocaInst>(StaleCI->getArgOperand(0));
1380       assert(ArgStructAlloca &&
1381              "Unable to find the alloca instruction corresponding to arguments "
1382              "for extracted function");
1383       StructType *ArgStructType =
1384           dyn_cast<StructType>(ArgStructAlloca->getAllocatedType());
1385       assert(ArgStructType && "Unable to find struct type corresponding to "
1386                               "arguments for extracted function");
1387       TaskSize =
1388           Builder.getInt64(M.getDataLayout().getTypeStoreSize(ArgStructType));
1389     }
1390 
1391     // TODO: Argument - sizeof_shareds
1392 
1393     // Argument - task_entry (the wrapper function)
1394     // If the outlined function has some captured variables (i.e. HasTaskData is
1395     // true), then the wrapper function will have an additional argument (the
1396     // struct containing captured variables). Otherwise, no such argument will
1397     // be present.
1398     SmallVector<Type *> WrapperArgTys{Builder.getInt32Ty()};
1399     if (HasTaskData)
1400       WrapperArgTys.push_back(OutlinedFn.getArg(0)->getType());
1401     FunctionCallee WrapperFuncVal = M.getOrInsertFunction(
1402         (Twine(OutlinedFn.getName()) + ".wrapper").str(),
1403         FunctionType::get(Builder.getInt32Ty(), WrapperArgTys, false));
1404     Function *WrapperFunc = dyn_cast<Function>(WrapperFuncVal.getCallee());
1405     PointerType *WrapperFuncBitcastType =
1406         FunctionType::get(Builder.getInt32Ty(),
1407                           {Builder.getInt32Ty(), Builder.getInt8PtrTy()}, false)
1408             ->getPointerTo();
1409     Value *WrapperFuncBitcast =
1410         ConstantExpr::getBitCast(WrapperFunc, WrapperFuncBitcastType);
1411 
1412     // Emit the @__kmpc_omp_task_alloc runtime call
1413     // The runtime call returns a pointer to an area where the task captured
1414     // variables must be copied before the task is run (NewTaskData)
1415     CallInst *NewTaskData = Builder.CreateCall(
1416         TaskAllocFn,
1417         {/*loc_ref=*/Ident, /*gtid=*/ThreadID, /*flags=*/Flags,
1418          /*sizeof_task=*/TaskSize, /*sizeof_shared=*/Builder.getInt64(0),
1419          /*task_func=*/WrapperFuncBitcast});
1420 
1421     // Copy the arguments for outlined function
1422     if (HasTaskData) {
1423       Value *TaskData = StaleCI->getArgOperand(0);
1424       Align Alignment = TaskData->getPointerAlignment(M.getDataLayout());
1425       Builder.CreateMemCpy(NewTaskData, Alignment, TaskData, Alignment,
1426                            TaskSize);
1427     }
1428 
1429     Value *DepArrayPtr = nullptr;
1430     if (Dependencies.size()) {
1431       InsertPointTy OldIP = Builder.saveIP();
1432       Builder.SetInsertPoint(
1433           &OldIP.getBlock()->getParent()->getEntryBlock().back());
1434 
1435       Type *DepArrayTy = ArrayType::get(DependInfo, Dependencies.size());
1436       Value *DepArray =
1437           Builder.CreateAlloca(DepArrayTy, nullptr, ".dep.arr.addr");
1438 
1439       unsigned P = 0;
1440       for (const DependData &Dep : Dependencies) {
1441         Value *Base =
1442             Builder.CreateConstInBoundsGEP2_64(DepArrayTy, DepArray, 0, P);
1443         // Store the pointer to the variable
1444         Value *Addr = Builder.CreateStructGEP(
1445             DependInfo, Base,
1446             static_cast<unsigned int>(RTLDependInfoFields::BaseAddr));
1447         Value *DepValPtr =
1448             Builder.CreatePtrToInt(Dep.DepVal, Builder.getInt64Ty());
1449         Builder.CreateStore(DepValPtr, Addr);
1450         // Store the size of the variable
1451         Value *Size = Builder.CreateStructGEP(
1452             DependInfo, Base,
1453             static_cast<unsigned int>(RTLDependInfoFields::Len));
1454         Builder.CreateStore(Builder.getInt64(M.getDataLayout().getTypeStoreSize(
1455                                 Dep.DepValueType)),
1456                             Size);
1457         // Store the dependency kind
1458         Value *Flags = Builder.CreateStructGEP(
1459             DependInfo, Base,
1460             static_cast<unsigned int>(RTLDependInfoFields::Flags));
1461         Builder.CreateStore(
1462             ConstantInt::get(Builder.getInt8Ty(),
1463                              static_cast<unsigned int>(Dep.DepKind)),
1464             Flags);
1465         ++P;
1466       }
1467 
1468       DepArrayPtr = Builder.CreateBitCast(DepArray, Builder.getInt8PtrTy());
1469       Builder.restoreIP(OldIP);
1470     }
1471 
1472     // In the presence of the `if` clause, the following IR is generated:
1473     //    ...
1474     //    %data = call @__kmpc_omp_task_alloc(...)
1475     //    br i1 %if_condition, label %then, label %else
1476     //  then:
1477     //    call @__kmpc_omp_task(...)
1478     //    br label %exit
1479     //  else:
1480     //    call @__kmpc_omp_task_begin_if0(...)
1481     //    call @wrapper_fn(...)
1482     //    call @__kmpc_omp_task_complete_if0(...)
1483     //    br label %exit
1484     //  exit:
1485     //    ...
1486     if (IfCondition) {
1487       // `SplitBlockAndInsertIfThenElse` requires the block to have a
1488       // terminator.
1489       BasicBlock *NewBasicBlock =
1490           splitBB(Builder, /*CreateBranch=*/true, "if.end");
1491       Instruction *IfTerminator =
1492           NewBasicBlock->getSinglePredecessor()->getTerminator();
1493       Instruction *ThenTI = IfTerminator, *ElseTI = nullptr;
1494       Builder.SetInsertPoint(IfTerminator);
1495       SplitBlockAndInsertIfThenElse(IfCondition, IfTerminator, &ThenTI,
1496                                     &ElseTI);
1497       Builder.SetInsertPoint(ElseTI);
1498       Function *TaskBeginFn =
1499           getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task_begin_if0);
1500       Function *TaskCompleteFn =
1501           getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task_complete_if0);
1502       Builder.CreateCall(TaskBeginFn, {Ident, ThreadID, NewTaskData});
1503       if (HasTaskData)
1504         Builder.CreateCall(WrapperFunc, {ThreadID, NewTaskData});
1505       else
1506         Builder.CreateCall(WrapperFunc, {ThreadID});
1507       Builder.CreateCall(TaskCompleteFn, {Ident, ThreadID, NewTaskData});
1508       Builder.SetInsertPoint(ThenTI);
1509     }
1510 
1511     if (Dependencies.size()) {
1512       Function *TaskFn =
1513           getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task_with_deps);
1514       Builder.CreateCall(
1515           TaskFn,
1516           {Ident, ThreadID, NewTaskData, Builder.getInt32(Dependencies.size()),
1517            DepArrayPtr, ConstantInt::get(Builder.getInt32Ty(), 0),
1518            ConstantPointerNull::get(Type::getInt8PtrTy(M.getContext()))});
1519 
1520     } else {
1521       // Emit the @__kmpc_omp_task runtime call to spawn the task
1522       Function *TaskFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task);
1523       Builder.CreateCall(TaskFn, {Ident, ThreadID, NewTaskData});
1524     }
1525 
1526     StaleCI->eraseFromParent();
1527 
1528     // Emit the body for wrapper function
1529     BasicBlock *WrapperEntryBB =
1530         BasicBlock::Create(M.getContext(), "", WrapperFunc);
1531     Builder.SetInsertPoint(WrapperEntryBB);
1532     if (HasTaskData)
1533       Builder.CreateCall(&OutlinedFn, {WrapperFunc->getArg(1)});
1534     else
1535       Builder.CreateCall(&OutlinedFn);
1536     Builder.CreateRet(Builder.getInt32(0));
1537   };
1538 
1539   addOutlineInfo(std::move(OI));
1540 
1541   InsertPointTy TaskAllocaIP =
1542       InsertPointTy(TaskAllocaBB, TaskAllocaBB->begin());
1543   InsertPointTy TaskBodyIP = InsertPointTy(TaskBodyBB, TaskBodyBB->begin());
1544   BodyGenCB(TaskAllocaIP, TaskBodyIP);
1545   Builder.SetInsertPoint(TaskExitBB, TaskExitBB->begin());
1546 
1547   return Builder.saveIP();
1548 }
1549 
1550 OpenMPIRBuilder::InsertPointTy
1551 OpenMPIRBuilder::createTaskgroup(const LocationDescription &Loc,
1552                                  InsertPointTy AllocaIP,
1553                                  BodyGenCallbackTy BodyGenCB) {
1554   if (!updateToLocation(Loc))
1555     return InsertPointTy();
1556 
1557   uint32_t SrcLocStrSize;
1558   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
1559   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
1560   Value *ThreadID = getOrCreateThreadID(Ident);
1561 
1562   // Emit the @__kmpc_taskgroup runtime call to start the taskgroup
1563   Function *TaskgroupFn =
1564       getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_taskgroup);
1565   Builder.CreateCall(TaskgroupFn, {Ident, ThreadID});
1566 
1567   BasicBlock *TaskgroupExitBB = splitBB(Builder, true, "taskgroup.exit");
1568   BodyGenCB(AllocaIP, Builder.saveIP());
1569 
1570   Builder.SetInsertPoint(TaskgroupExitBB);
1571   // Emit the @__kmpc_end_taskgroup runtime call to end the taskgroup
1572   Function *EndTaskgroupFn =
1573       getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_taskgroup);
1574   Builder.CreateCall(EndTaskgroupFn, {Ident, ThreadID});
1575 
1576   return Builder.saveIP();
1577 }
1578 
1579 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createSections(
1580     const LocationDescription &Loc, InsertPointTy AllocaIP,
1581     ArrayRef<StorableBodyGenCallbackTy> SectionCBs, PrivatizeCallbackTy PrivCB,
1582     FinalizeCallbackTy FiniCB, bool IsCancellable, bool IsNowait) {
1583   assert(!isConflictIP(AllocaIP, Loc.IP) && "Dedicated IP allocas required");
1584 
1585   if (!updateToLocation(Loc))
1586     return Loc.IP;
1587 
1588   auto FiniCBWrapper = [&](InsertPointTy IP) {
1589     if (IP.getBlock()->end() != IP.getPoint())
1590       return FiniCB(IP);
1591     // This must be done otherwise any nested constructs using FinalizeOMPRegion
1592     // will fail because that function requires the Finalization Basic Block to
1593     // have a terminator, which is already removed by EmitOMPRegionBody.
1594     // IP is currently at cancelation block.
1595     // We need to backtrack to the condition block to fetch
1596     // the exit block and create a branch from cancelation
1597     // to exit block.
1598     IRBuilder<>::InsertPointGuard IPG(Builder);
1599     Builder.restoreIP(IP);
1600     auto *CaseBB = IP.getBlock()->getSinglePredecessor();
1601     auto *CondBB = CaseBB->getSinglePredecessor()->getSinglePredecessor();
1602     auto *ExitBB = CondBB->getTerminator()->getSuccessor(1);
1603     Instruction *I = Builder.CreateBr(ExitBB);
1604     IP = InsertPointTy(I->getParent(), I->getIterator());
1605     return FiniCB(IP);
1606   };
1607 
1608   FinalizationStack.push_back({FiniCBWrapper, OMPD_sections, IsCancellable});
1609 
1610   // Each section is emitted as a switch case
1611   // Each finalization callback is handled from clang.EmitOMPSectionDirective()
1612   // -> OMP.createSection() which generates the IR for each section
1613   // Iterate through all sections and emit a switch construct:
1614   // switch (IV) {
1615   //   case 0:
1616   //     <SectionStmt[0]>;
1617   //     break;
1618   // ...
1619   //   case <NumSection> - 1:
1620   //     <SectionStmt[<NumSection> - 1]>;
1621   //     break;
1622   // }
1623   // ...
1624   // section_loop.after:
1625   // <FiniCB>;
1626   auto LoopBodyGenCB = [&](InsertPointTy CodeGenIP, Value *IndVar) {
1627     Builder.restoreIP(CodeGenIP);
1628     BasicBlock *Continue =
1629         splitBBWithSuffix(Builder, /*CreateBranch=*/false, ".sections.after");
1630     Function *CurFn = Continue->getParent();
1631     SwitchInst *SwitchStmt = Builder.CreateSwitch(IndVar, Continue);
1632 
1633     unsigned CaseNumber = 0;
1634     for (auto SectionCB : SectionCBs) {
1635       BasicBlock *CaseBB = BasicBlock::Create(
1636           M.getContext(), "omp_section_loop.body.case", CurFn, Continue);
1637       SwitchStmt->addCase(Builder.getInt32(CaseNumber), CaseBB);
1638       Builder.SetInsertPoint(CaseBB);
1639       BranchInst *CaseEndBr = Builder.CreateBr(Continue);
1640       SectionCB(InsertPointTy(),
1641                 {CaseEndBr->getParent(), CaseEndBr->getIterator()});
1642       CaseNumber++;
1643     }
1644     // remove the existing terminator from body BB since there can be no
1645     // terminators after switch/case
1646   };
1647   // Loop body ends here
1648   // LowerBound, UpperBound, and STride for createCanonicalLoop
1649   Type *I32Ty = Type::getInt32Ty(M.getContext());
1650   Value *LB = ConstantInt::get(I32Ty, 0);
1651   Value *UB = ConstantInt::get(I32Ty, SectionCBs.size());
1652   Value *ST = ConstantInt::get(I32Ty, 1);
1653   llvm::CanonicalLoopInfo *LoopInfo = createCanonicalLoop(
1654       Loc, LoopBodyGenCB, LB, UB, ST, true, false, AllocaIP, "section_loop");
1655   InsertPointTy AfterIP =
1656       applyStaticWorkshareLoop(Loc.DL, LoopInfo, AllocaIP, !IsNowait);
1657 
1658   // Apply the finalization callback in LoopAfterBB
1659   auto FiniInfo = FinalizationStack.pop_back_val();
1660   assert(FiniInfo.DK == OMPD_sections &&
1661          "Unexpected finalization stack state!");
1662   if (FinalizeCallbackTy &CB = FiniInfo.FiniCB) {
1663     Builder.restoreIP(AfterIP);
1664     BasicBlock *FiniBB =
1665         splitBBWithSuffix(Builder, /*CreateBranch=*/true, "sections.fini");
1666     CB(Builder.saveIP());
1667     AfterIP = {FiniBB, FiniBB->begin()};
1668   }
1669 
1670   return AfterIP;
1671 }
1672 
1673 OpenMPIRBuilder::InsertPointTy
1674 OpenMPIRBuilder::createSection(const LocationDescription &Loc,
1675                                BodyGenCallbackTy BodyGenCB,
1676                                FinalizeCallbackTy FiniCB) {
1677   if (!updateToLocation(Loc))
1678     return Loc.IP;
1679 
1680   auto FiniCBWrapper = [&](InsertPointTy IP) {
1681     if (IP.getBlock()->end() != IP.getPoint())
1682       return FiniCB(IP);
1683     // This must be done otherwise any nested constructs using FinalizeOMPRegion
1684     // will fail because that function requires the Finalization Basic Block to
1685     // have a terminator, which is already removed by EmitOMPRegionBody.
1686     // IP is currently at cancelation block.
1687     // We need to backtrack to the condition block to fetch
1688     // the exit block and create a branch from cancelation
1689     // to exit block.
1690     IRBuilder<>::InsertPointGuard IPG(Builder);
1691     Builder.restoreIP(IP);
1692     auto *CaseBB = Loc.IP.getBlock();
1693     auto *CondBB = CaseBB->getSinglePredecessor()->getSinglePredecessor();
1694     auto *ExitBB = CondBB->getTerminator()->getSuccessor(1);
1695     Instruction *I = Builder.CreateBr(ExitBB);
1696     IP = InsertPointTy(I->getParent(), I->getIterator());
1697     return FiniCB(IP);
1698   };
1699 
1700   Directive OMPD = Directive::OMPD_sections;
1701   // Since we are using Finalization Callback here, HasFinalize
1702   // and IsCancellable have to be true
1703   return EmitOMPInlinedRegion(OMPD, nullptr, nullptr, BodyGenCB, FiniCBWrapper,
1704                               /*Conditional*/ false, /*hasFinalize*/ true,
1705                               /*IsCancellable*/ true);
1706 }
1707 
1708 /// Create a function with a unique name and a "void (i8*, i8*)" signature in
1709 /// the given module and return it.
1710 Function *getFreshReductionFunc(Module &M) {
1711   Type *VoidTy = Type::getVoidTy(M.getContext());
1712   Type *Int8PtrTy = Type::getInt8PtrTy(M.getContext());
1713   auto *FuncTy =
1714       FunctionType::get(VoidTy, {Int8PtrTy, Int8PtrTy}, /* IsVarArg */ false);
1715   return Function::Create(FuncTy, GlobalVariable::InternalLinkage,
1716                           M.getDataLayout().getDefaultGlobalsAddressSpace(),
1717                           ".omp.reduction.func", &M);
1718 }
1719 
1720 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createReductions(
1721     const LocationDescription &Loc, InsertPointTy AllocaIP,
1722     ArrayRef<ReductionInfo> ReductionInfos, bool IsNoWait) {
1723   for (const ReductionInfo &RI : ReductionInfos) {
1724     (void)RI;
1725     assert(RI.Variable && "expected non-null variable");
1726     assert(RI.PrivateVariable && "expected non-null private variable");
1727     assert(RI.ReductionGen && "expected non-null reduction generator callback");
1728     assert(RI.Variable->getType() == RI.PrivateVariable->getType() &&
1729            "expected variables and their private equivalents to have the same "
1730            "type");
1731     assert(RI.Variable->getType()->isPointerTy() &&
1732            "expected variables to be pointers");
1733   }
1734 
1735   if (!updateToLocation(Loc))
1736     return InsertPointTy();
1737 
1738   BasicBlock *InsertBlock = Loc.IP.getBlock();
1739   BasicBlock *ContinuationBlock =
1740       InsertBlock->splitBasicBlock(Loc.IP.getPoint(), "reduce.finalize");
1741   InsertBlock->getTerminator()->eraseFromParent();
1742 
1743   // Create and populate array of type-erased pointers to private reduction
1744   // values.
1745   unsigned NumReductions = ReductionInfos.size();
1746   Type *RedArrayTy = ArrayType::get(Builder.getInt8PtrTy(), NumReductions);
1747   Builder.restoreIP(AllocaIP);
1748   Value *RedArray = Builder.CreateAlloca(RedArrayTy, nullptr, "red.array");
1749 
1750   Builder.SetInsertPoint(InsertBlock, InsertBlock->end());
1751 
1752   for (auto En : enumerate(ReductionInfos)) {
1753     unsigned Index = En.index();
1754     const ReductionInfo &RI = En.value();
1755     Value *RedArrayElemPtr = Builder.CreateConstInBoundsGEP2_64(
1756         RedArrayTy, RedArray, 0, Index, "red.array.elem." + Twine(Index));
1757     Value *Casted =
1758         Builder.CreateBitCast(RI.PrivateVariable, Builder.getInt8PtrTy(),
1759                               "private.red.var." + Twine(Index) + ".casted");
1760     Builder.CreateStore(Casted, RedArrayElemPtr);
1761   }
1762 
1763   // Emit a call to the runtime function that orchestrates the reduction.
1764   // Declare the reduction function in the process.
1765   Function *Func = Builder.GetInsertBlock()->getParent();
1766   Module *Module = Func->getParent();
1767   Value *RedArrayPtr =
1768       Builder.CreateBitCast(RedArray, Builder.getInt8PtrTy(), "red.array.ptr");
1769   uint32_t SrcLocStrSize;
1770   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
1771   bool CanGenerateAtomic =
1772       llvm::all_of(ReductionInfos, [](const ReductionInfo &RI) {
1773         return RI.AtomicReductionGen;
1774       });
1775   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize,
1776                                   CanGenerateAtomic
1777                                       ? IdentFlag::OMP_IDENT_FLAG_ATOMIC_REDUCE
1778                                       : IdentFlag(0));
1779   Value *ThreadId = getOrCreateThreadID(Ident);
1780   Constant *NumVariables = Builder.getInt32(NumReductions);
1781   const DataLayout &DL = Module->getDataLayout();
1782   unsigned RedArrayByteSize = DL.getTypeStoreSize(RedArrayTy);
1783   Constant *RedArraySize = Builder.getInt64(RedArrayByteSize);
1784   Function *ReductionFunc = getFreshReductionFunc(*Module);
1785   Value *Lock = getOMPCriticalRegionLock(".reduction");
1786   Function *ReduceFunc = getOrCreateRuntimeFunctionPtr(
1787       IsNoWait ? RuntimeFunction::OMPRTL___kmpc_reduce_nowait
1788                : RuntimeFunction::OMPRTL___kmpc_reduce);
1789   CallInst *ReduceCall =
1790       Builder.CreateCall(ReduceFunc,
1791                          {Ident, ThreadId, NumVariables, RedArraySize,
1792                           RedArrayPtr, ReductionFunc, Lock},
1793                          "reduce");
1794 
1795   // Create final reduction entry blocks for the atomic and non-atomic case.
1796   // Emit IR that dispatches control flow to one of the blocks based on the
1797   // reduction supporting the atomic mode.
1798   BasicBlock *NonAtomicRedBlock =
1799       BasicBlock::Create(Module->getContext(), "reduce.switch.nonatomic", Func);
1800   BasicBlock *AtomicRedBlock =
1801       BasicBlock::Create(Module->getContext(), "reduce.switch.atomic", Func);
1802   SwitchInst *Switch =
1803       Builder.CreateSwitch(ReduceCall, ContinuationBlock, /* NumCases */ 2);
1804   Switch->addCase(Builder.getInt32(1), NonAtomicRedBlock);
1805   Switch->addCase(Builder.getInt32(2), AtomicRedBlock);
1806 
1807   // Populate the non-atomic reduction using the elementwise reduction function.
1808   // This loads the elements from the global and private variables and reduces
1809   // them before storing back the result to the global variable.
1810   Builder.SetInsertPoint(NonAtomicRedBlock);
1811   for (auto En : enumerate(ReductionInfos)) {
1812     const ReductionInfo &RI = En.value();
1813     Type *ValueType = RI.ElementType;
1814     Value *RedValue = Builder.CreateLoad(ValueType, RI.Variable,
1815                                          "red.value." + Twine(En.index()));
1816     Value *PrivateRedValue =
1817         Builder.CreateLoad(ValueType, RI.PrivateVariable,
1818                            "red.private.value." + Twine(En.index()));
1819     Value *Reduced;
1820     Builder.restoreIP(
1821         RI.ReductionGen(Builder.saveIP(), RedValue, PrivateRedValue, Reduced));
1822     if (!Builder.GetInsertBlock())
1823       return InsertPointTy();
1824     Builder.CreateStore(Reduced, RI.Variable);
1825   }
1826   Function *EndReduceFunc = getOrCreateRuntimeFunctionPtr(
1827       IsNoWait ? RuntimeFunction::OMPRTL___kmpc_end_reduce_nowait
1828                : RuntimeFunction::OMPRTL___kmpc_end_reduce);
1829   Builder.CreateCall(EndReduceFunc, {Ident, ThreadId, Lock});
1830   Builder.CreateBr(ContinuationBlock);
1831 
1832   // Populate the atomic reduction using the atomic elementwise reduction
1833   // function. There are no loads/stores here because they will be happening
1834   // inside the atomic elementwise reduction.
1835   Builder.SetInsertPoint(AtomicRedBlock);
1836   if (CanGenerateAtomic) {
1837     for (const ReductionInfo &RI : ReductionInfos) {
1838       Builder.restoreIP(RI.AtomicReductionGen(Builder.saveIP(), RI.ElementType,
1839                                               RI.Variable, RI.PrivateVariable));
1840       if (!Builder.GetInsertBlock())
1841         return InsertPointTy();
1842     }
1843     Builder.CreateBr(ContinuationBlock);
1844   } else {
1845     Builder.CreateUnreachable();
1846   }
1847 
1848   // Populate the outlined reduction function using the elementwise reduction
1849   // function. Partial values are extracted from the type-erased array of
1850   // pointers to private variables.
1851   BasicBlock *ReductionFuncBlock =
1852       BasicBlock::Create(Module->getContext(), "", ReductionFunc);
1853   Builder.SetInsertPoint(ReductionFuncBlock);
1854   Value *LHSArrayPtr = Builder.CreateBitCast(ReductionFunc->getArg(0),
1855                                              RedArrayTy->getPointerTo());
1856   Value *RHSArrayPtr = Builder.CreateBitCast(ReductionFunc->getArg(1),
1857                                              RedArrayTy->getPointerTo());
1858   for (auto En : enumerate(ReductionInfos)) {
1859     const ReductionInfo &RI = En.value();
1860     Value *LHSI8PtrPtr = Builder.CreateConstInBoundsGEP2_64(
1861         RedArrayTy, LHSArrayPtr, 0, En.index());
1862     Value *LHSI8Ptr = Builder.CreateLoad(Builder.getInt8PtrTy(), LHSI8PtrPtr);
1863     Value *LHSPtr = Builder.CreateBitCast(LHSI8Ptr, RI.Variable->getType());
1864     Value *LHS = Builder.CreateLoad(RI.ElementType, LHSPtr);
1865     Value *RHSI8PtrPtr = Builder.CreateConstInBoundsGEP2_64(
1866         RedArrayTy, RHSArrayPtr, 0, En.index());
1867     Value *RHSI8Ptr = Builder.CreateLoad(Builder.getInt8PtrTy(), RHSI8PtrPtr);
1868     Value *RHSPtr =
1869         Builder.CreateBitCast(RHSI8Ptr, RI.PrivateVariable->getType());
1870     Value *RHS = Builder.CreateLoad(RI.ElementType, RHSPtr);
1871     Value *Reduced;
1872     Builder.restoreIP(RI.ReductionGen(Builder.saveIP(), LHS, RHS, Reduced));
1873     if (!Builder.GetInsertBlock())
1874       return InsertPointTy();
1875     Builder.CreateStore(Reduced, LHSPtr);
1876   }
1877   Builder.CreateRetVoid();
1878 
1879   Builder.SetInsertPoint(ContinuationBlock);
1880   return Builder.saveIP();
1881 }
1882 
1883 OpenMPIRBuilder::InsertPointTy
1884 OpenMPIRBuilder::createMaster(const LocationDescription &Loc,
1885                               BodyGenCallbackTy BodyGenCB,
1886                               FinalizeCallbackTy FiniCB) {
1887 
1888   if (!updateToLocation(Loc))
1889     return Loc.IP;
1890 
1891   Directive OMPD = Directive::OMPD_master;
1892   uint32_t SrcLocStrSize;
1893   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
1894   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
1895   Value *ThreadId = getOrCreateThreadID(Ident);
1896   Value *Args[] = {Ident, ThreadId};
1897 
1898   Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_master);
1899   Instruction *EntryCall = Builder.CreateCall(EntryRTLFn, Args);
1900 
1901   Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_master);
1902   Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, Args);
1903 
1904   return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
1905                               /*Conditional*/ true, /*hasFinalize*/ true);
1906 }
1907 
1908 OpenMPIRBuilder::InsertPointTy
1909 OpenMPIRBuilder::createMasked(const LocationDescription &Loc,
1910                               BodyGenCallbackTy BodyGenCB,
1911                               FinalizeCallbackTy FiniCB, Value *Filter) {
1912   if (!updateToLocation(Loc))
1913     return Loc.IP;
1914 
1915   Directive OMPD = Directive::OMPD_masked;
1916   uint32_t SrcLocStrSize;
1917   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
1918   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
1919   Value *ThreadId = getOrCreateThreadID(Ident);
1920   Value *Args[] = {Ident, ThreadId, Filter};
1921   Value *ArgsEnd[] = {Ident, ThreadId};
1922 
1923   Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_masked);
1924   Instruction *EntryCall = Builder.CreateCall(EntryRTLFn, Args);
1925 
1926   Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_masked);
1927   Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, ArgsEnd);
1928 
1929   return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
1930                               /*Conditional*/ true, /*hasFinalize*/ true);
1931 }
1932 
1933 CanonicalLoopInfo *OpenMPIRBuilder::createLoopSkeleton(
1934     DebugLoc DL, Value *TripCount, Function *F, BasicBlock *PreInsertBefore,
1935     BasicBlock *PostInsertBefore, const Twine &Name) {
1936   Module *M = F->getParent();
1937   LLVMContext &Ctx = M->getContext();
1938   Type *IndVarTy = TripCount->getType();
1939 
1940   // Create the basic block structure.
1941   BasicBlock *Preheader =
1942       BasicBlock::Create(Ctx, "omp_" + Name + ".preheader", F, PreInsertBefore);
1943   BasicBlock *Header =
1944       BasicBlock::Create(Ctx, "omp_" + Name + ".header", F, PreInsertBefore);
1945   BasicBlock *Cond =
1946       BasicBlock::Create(Ctx, "omp_" + Name + ".cond", F, PreInsertBefore);
1947   BasicBlock *Body =
1948       BasicBlock::Create(Ctx, "omp_" + Name + ".body", F, PreInsertBefore);
1949   BasicBlock *Latch =
1950       BasicBlock::Create(Ctx, "omp_" + Name + ".inc", F, PostInsertBefore);
1951   BasicBlock *Exit =
1952       BasicBlock::Create(Ctx, "omp_" + Name + ".exit", F, PostInsertBefore);
1953   BasicBlock *After =
1954       BasicBlock::Create(Ctx, "omp_" + Name + ".after", F, PostInsertBefore);
1955 
1956   // Use specified DebugLoc for new instructions.
1957   Builder.SetCurrentDebugLocation(DL);
1958 
1959   Builder.SetInsertPoint(Preheader);
1960   Builder.CreateBr(Header);
1961 
1962   Builder.SetInsertPoint(Header);
1963   PHINode *IndVarPHI = Builder.CreatePHI(IndVarTy, 2, "omp_" + Name + ".iv");
1964   IndVarPHI->addIncoming(ConstantInt::get(IndVarTy, 0), Preheader);
1965   Builder.CreateBr(Cond);
1966 
1967   Builder.SetInsertPoint(Cond);
1968   Value *Cmp =
1969       Builder.CreateICmpULT(IndVarPHI, TripCount, "omp_" + Name + ".cmp");
1970   Builder.CreateCondBr(Cmp, Body, Exit);
1971 
1972   Builder.SetInsertPoint(Body);
1973   Builder.CreateBr(Latch);
1974 
1975   Builder.SetInsertPoint(Latch);
1976   Value *Next = Builder.CreateAdd(IndVarPHI, ConstantInt::get(IndVarTy, 1),
1977                                   "omp_" + Name + ".next", /*HasNUW=*/true);
1978   Builder.CreateBr(Header);
1979   IndVarPHI->addIncoming(Next, Latch);
1980 
1981   Builder.SetInsertPoint(Exit);
1982   Builder.CreateBr(After);
1983 
1984   // Remember and return the canonical control flow.
1985   LoopInfos.emplace_front();
1986   CanonicalLoopInfo *CL = &LoopInfos.front();
1987 
1988   CL->Header = Header;
1989   CL->Cond = Cond;
1990   CL->Latch = Latch;
1991   CL->Exit = Exit;
1992 
1993 #ifndef NDEBUG
1994   CL->assertOK();
1995 #endif
1996   return CL;
1997 }
1998 
1999 CanonicalLoopInfo *
2000 OpenMPIRBuilder::createCanonicalLoop(const LocationDescription &Loc,
2001                                      LoopBodyGenCallbackTy BodyGenCB,
2002                                      Value *TripCount, const Twine &Name) {
2003   BasicBlock *BB = Loc.IP.getBlock();
2004   BasicBlock *NextBB = BB->getNextNode();
2005 
2006   CanonicalLoopInfo *CL = createLoopSkeleton(Loc.DL, TripCount, BB->getParent(),
2007                                              NextBB, NextBB, Name);
2008   BasicBlock *After = CL->getAfter();
2009 
2010   // If location is not set, don't connect the loop.
2011   if (updateToLocation(Loc)) {
2012     // Split the loop at the insertion point: Branch to the preheader and move
2013     // every following instruction to after the loop (the After BB). Also, the
2014     // new successor is the loop's after block.
2015     spliceBB(Builder, After, /*CreateBranch=*/false);
2016     Builder.CreateBr(CL->getPreheader());
2017   }
2018 
2019   // Emit the body content. We do it after connecting the loop to the CFG to
2020   // avoid that the callback encounters degenerate BBs.
2021   BodyGenCB(CL->getBodyIP(), CL->getIndVar());
2022 
2023 #ifndef NDEBUG
2024   CL->assertOK();
2025 #endif
2026   return CL;
2027 }
2028 
2029 CanonicalLoopInfo *OpenMPIRBuilder::createCanonicalLoop(
2030     const LocationDescription &Loc, LoopBodyGenCallbackTy BodyGenCB,
2031     Value *Start, Value *Stop, Value *Step, bool IsSigned, bool InclusiveStop,
2032     InsertPointTy ComputeIP, const Twine &Name) {
2033 
2034   // Consider the following difficulties (assuming 8-bit signed integers):
2035   //  * Adding \p Step to the loop counter which passes \p Stop may overflow:
2036   //      DO I = 1, 100, 50
2037   ///  * A \p Step of INT_MIN cannot not be normalized to a positive direction:
2038   //      DO I = 100, 0, -128
2039 
2040   // Start, Stop and Step must be of the same integer type.
2041   auto *IndVarTy = cast<IntegerType>(Start->getType());
2042   assert(IndVarTy == Stop->getType() && "Stop type mismatch");
2043   assert(IndVarTy == Step->getType() && "Step type mismatch");
2044 
2045   LocationDescription ComputeLoc =
2046       ComputeIP.isSet() ? LocationDescription(ComputeIP, Loc.DL) : Loc;
2047   updateToLocation(ComputeLoc);
2048 
2049   ConstantInt *Zero = ConstantInt::get(IndVarTy, 0);
2050   ConstantInt *One = ConstantInt::get(IndVarTy, 1);
2051 
2052   // Like Step, but always positive.
2053   Value *Incr = Step;
2054 
2055   // Distance between Start and Stop; always positive.
2056   Value *Span;
2057 
2058   // Condition whether there are no iterations are executed at all, e.g. because
2059   // UB < LB.
2060   Value *ZeroCmp;
2061 
2062   if (IsSigned) {
2063     // Ensure that increment is positive. If not, negate and invert LB and UB.
2064     Value *IsNeg = Builder.CreateICmpSLT(Step, Zero);
2065     Incr = Builder.CreateSelect(IsNeg, Builder.CreateNeg(Step), Step);
2066     Value *LB = Builder.CreateSelect(IsNeg, Stop, Start);
2067     Value *UB = Builder.CreateSelect(IsNeg, Start, Stop);
2068     Span = Builder.CreateSub(UB, LB, "", false, true);
2069     ZeroCmp = Builder.CreateICmp(
2070         InclusiveStop ? CmpInst::ICMP_SLT : CmpInst::ICMP_SLE, UB, LB);
2071   } else {
2072     Span = Builder.CreateSub(Stop, Start, "", true);
2073     ZeroCmp = Builder.CreateICmp(
2074         InclusiveStop ? CmpInst::ICMP_ULT : CmpInst::ICMP_ULE, Stop, Start);
2075   }
2076 
2077   Value *CountIfLooping;
2078   if (InclusiveStop) {
2079     CountIfLooping = Builder.CreateAdd(Builder.CreateUDiv(Span, Incr), One);
2080   } else {
2081     // Avoid incrementing past stop since it could overflow.
2082     Value *CountIfTwo = Builder.CreateAdd(
2083         Builder.CreateUDiv(Builder.CreateSub(Span, One), Incr), One);
2084     Value *OneCmp = Builder.CreateICmp(
2085         InclusiveStop ? CmpInst::ICMP_ULT : CmpInst::ICMP_ULE, Span, Incr);
2086     CountIfLooping = Builder.CreateSelect(OneCmp, One, CountIfTwo);
2087   }
2088   Value *TripCount = Builder.CreateSelect(ZeroCmp, Zero, CountIfLooping,
2089                                           "omp_" + Name + ".tripcount");
2090 
2091   auto BodyGen = [=](InsertPointTy CodeGenIP, Value *IV) {
2092     Builder.restoreIP(CodeGenIP);
2093     Value *Span = Builder.CreateMul(IV, Step);
2094     Value *IndVar = Builder.CreateAdd(Span, Start);
2095     BodyGenCB(Builder.saveIP(), IndVar);
2096   };
2097   LocationDescription LoopLoc = ComputeIP.isSet() ? Loc.IP : Builder.saveIP();
2098   return createCanonicalLoop(LoopLoc, BodyGen, TripCount, Name);
2099 }
2100 
2101 // Returns an LLVM function to call for initializing loop bounds using OpenMP
2102 // static scheduling depending on `type`. Only i32 and i64 are supported by the
2103 // runtime. Always interpret integers as unsigned similarly to
2104 // CanonicalLoopInfo.
2105 static FunctionCallee getKmpcForStaticInitForType(Type *Ty, Module &M,
2106                                                   OpenMPIRBuilder &OMPBuilder) {
2107   unsigned Bitwidth = Ty->getIntegerBitWidth();
2108   if (Bitwidth == 32)
2109     return OMPBuilder.getOrCreateRuntimeFunction(
2110         M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_4u);
2111   if (Bitwidth == 64)
2112     return OMPBuilder.getOrCreateRuntimeFunction(
2113         M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_8u);
2114   llvm_unreachable("unknown OpenMP loop iterator bitwidth");
2115 }
2116 
2117 OpenMPIRBuilder::InsertPointTy
2118 OpenMPIRBuilder::applyStaticWorkshareLoop(DebugLoc DL, CanonicalLoopInfo *CLI,
2119                                           InsertPointTy AllocaIP,
2120                                           bool NeedsBarrier) {
2121   assert(CLI->isValid() && "Requires a valid canonical loop");
2122   assert(!isConflictIP(AllocaIP, CLI->getPreheaderIP()) &&
2123          "Require dedicated allocate IP");
2124 
2125   // Set up the source location value for OpenMP runtime.
2126   Builder.restoreIP(CLI->getPreheaderIP());
2127   Builder.SetCurrentDebugLocation(DL);
2128 
2129   uint32_t SrcLocStrSize;
2130   Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize);
2131   Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
2132 
2133   // Declare useful OpenMP runtime functions.
2134   Value *IV = CLI->getIndVar();
2135   Type *IVTy = IV->getType();
2136   FunctionCallee StaticInit = getKmpcForStaticInitForType(IVTy, M, *this);
2137   FunctionCallee StaticFini =
2138       getOrCreateRuntimeFunction(M, omp::OMPRTL___kmpc_for_static_fini);
2139 
2140   // Allocate space for computed loop bounds as expected by the "init" function.
2141   Builder.restoreIP(AllocaIP);
2142   Type *I32Type = Type::getInt32Ty(M.getContext());
2143   Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter");
2144   Value *PLowerBound = Builder.CreateAlloca(IVTy, nullptr, "p.lowerbound");
2145   Value *PUpperBound = Builder.CreateAlloca(IVTy, nullptr, "p.upperbound");
2146   Value *PStride = Builder.CreateAlloca(IVTy, nullptr, "p.stride");
2147 
2148   // At the end of the preheader, prepare for calling the "init" function by
2149   // storing the current loop bounds into the allocated space. A canonical loop
2150   // always iterates from 0 to trip-count with step 1. Note that "init" expects
2151   // and produces an inclusive upper bound.
2152   Builder.SetInsertPoint(CLI->getPreheader()->getTerminator());
2153   Constant *Zero = ConstantInt::get(IVTy, 0);
2154   Constant *One = ConstantInt::get(IVTy, 1);
2155   Builder.CreateStore(Zero, PLowerBound);
2156   Value *UpperBound = Builder.CreateSub(CLI->getTripCount(), One);
2157   Builder.CreateStore(UpperBound, PUpperBound);
2158   Builder.CreateStore(One, PStride);
2159 
2160   Value *ThreadNum = getOrCreateThreadID(SrcLoc);
2161 
2162   Constant *SchedulingType = ConstantInt::get(
2163       I32Type, static_cast<int>(OMPScheduleType::UnorderedStatic));
2164 
2165   // Call the "init" function and update the trip count of the loop with the
2166   // value it produced.
2167   Builder.CreateCall(StaticInit,
2168                      {SrcLoc, ThreadNum, SchedulingType, PLastIter, PLowerBound,
2169                       PUpperBound, PStride, One, Zero});
2170   Value *LowerBound = Builder.CreateLoad(IVTy, PLowerBound);
2171   Value *InclusiveUpperBound = Builder.CreateLoad(IVTy, PUpperBound);
2172   Value *TripCountMinusOne = Builder.CreateSub(InclusiveUpperBound, LowerBound);
2173   Value *TripCount = Builder.CreateAdd(TripCountMinusOne, One);
2174   CLI->setTripCount(TripCount);
2175 
2176   // Update all uses of the induction variable except the one in the condition
2177   // block that compares it with the actual upper bound, and the increment in
2178   // the latch block.
2179 
2180   CLI->mapIndVar([&](Instruction *OldIV) -> Value * {
2181     Builder.SetInsertPoint(CLI->getBody(),
2182                            CLI->getBody()->getFirstInsertionPt());
2183     Builder.SetCurrentDebugLocation(DL);
2184     return Builder.CreateAdd(OldIV, LowerBound);
2185   });
2186 
2187   // In the "exit" block, call the "fini" function.
2188   Builder.SetInsertPoint(CLI->getExit(),
2189                          CLI->getExit()->getTerminator()->getIterator());
2190   Builder.CreateCall(StaticFini, {SrcLoc, ThreadNum});
2191 
2192   // Add the barrier if requested.
2193   if (NeedsBarrier)
2194     createBarrier(LocationDescription(Builder.saveIP(), DL),
2195                   omp::Directive::OMPD_for, /* ForceSimpleCall */ false,
2196                   /* CheckCancelFlag */ false);
2197 
2198   InsertPointTy AfterIP = CLI->getAfterIP();
2199   CLI->invalidate();
2200 
2201   return AfterIP;
2202 }
2203 
2204 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyStaticChunkedWorkshareLoop(
2205     DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP,
2206     bool NeedsBarrier, Value *ChunkSize) {
2207   assert(CLI->isValid() && "Requires a valid canonical loop");
2208   assert(ChunkSize && "Chunk size is required");
2209 
2210   LLVMContext &Ctx = CLI->getFunction()->getContext();
2211   Value *IV = CLI->getIndVar();
2212   Value *OrigTripCount = CLI->getTripCount();
2213   Type *IVTy = IV->getType();
2214   assert(IVTy->getIntegerBitWidth() <= 64 &&
2215          "Max supported tripcount bitwidth is 64 bits");
2216   Type *InternalIVTy = IVTy->getIntegerBitWidth() <= 32 ? Type::getInt32Ty(Ctx)
2217                                                         : Type::getInt64Ty(Ctx);
2218   Type *I32Type = Type::getInt32Ty(M.getContext());
2219   Constant *Zero = ConstantInt::get(InternalIVTy, 0);
2220   Constant *One = ConstantInt::get(InternalIVTy, 1);
2221 
2222   // Declare useful OpenMP runtime functions.
2223   FunctionCallee StaticInit =
2224       getKmpcForStaticInitForType(InternalIVTy, M, *this);
2225   FunctionCallee StaticFini =
2226       getOrCreateRuntimeFunction(M, omp::OMPRTL___kmpc_for_static_fini);
2227 
2228   // Allocate space for computed loop bounds as expected by the "init" function.
2229   Builder.restoreIP(AllocaIP);
2230   Builder.SetCurrentDebugLocation(DL);
2231   Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter");
2232   Value *PLowerBound =
2233       Builder.CreateAlloca(InternalIVTy, nullptr, "p.lowerbound");
2234   Value *PUpperBound =
2235       Builder.CreateAlloca(InternalIVTy, nullptr, "p.upperbound");
2236   Value *PStride = Builder.CreateAlloca(InternalIVTy, nullptr, "p.stride");
2237 
2238   // Set up the source location value for the OpenMP runtime.
2239   Builder.restoreIP(CLI->getPreheaderIP());
2240   Builder.SetCurrentDebugLocation(DL);
2241 
2242   // TODO: Detect overflow in ubsan or max-out with current tripcount.
2243   Value *CastedChunkSize =
2244       Builder.CreateZExtOrTrunc(ChunkSize, InternalIVTy, "chunksize");
2245   Value *CastedTripCount =
2246       Builder.CreateZExt(OrigTripCount, InternalIVTy, "tripcount");
2247 
2248   Constant *SchedulingType = ConstantInt::get(
2249       I32Type, static_cast<int>(OMPScheduleType::UnorderedStaticChunked));
2250   Builder.CreateStore(Zero, PLowerBound);
2251   Value *OrigUpperBound = Builder.CreateSub(CastedTripCount, One);
2252   Builder.CreateStore(OrigUpperBound, PUpperBound);
2253   Builder.CreateStore(One, PStride);
2254 
2255   // Call the "init" function and update the trip count of the loop with the
2256   // value it produced.
2257   uint32_t SrcLocStrSize;
2258   Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize);
2259   Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
2260   Value *ThreadNum = getOrCreateThreadID(SrcLoc);
2261   Builder.CreateCall(StaticInit,
2262                      {/*loc=*/SrcLoc, /*global_tid=*/ThreadNum,
2263                       /*schedtype=*/SchedulingType, /*plastiter=*/PLastIter,
2264                       /*plower=*/PLowerBound, /*pupper=*/PUpperBound,
2265                       /*pstride=*/PStride, /*incr=*/One,
2266                       /*chunk=*/CastedChunkSize});
2267 
2268   // Load values written by the "init" function.
2269   Value *FirstChunkStart =
2270       Builder.CreateLoad(InternalIVTy, PLowerBound, "omp_firstchunk.lb");
2271   Value *FirstChunkStop =
2272       Builder.CreateLoad(InternalIVTy, PUpperBound, "omp_firstchunk.ub");
2273   Value *FirstChunkEnd = Builder.CreateAdd(FirstChunkStop, One);
2274   Value *ChunkRange =
2275       Builder.CreateSub(FirstChunkEnd, FirstChunkStart, "omp_chunk.range");
2276   Value *NextChunkStride =
2277       Builder.CreateLoad(InternalIVTy, PStride, "omp_dispatch.stride");
2278 
2279   // Create outer "dispatch" loop for enumerating the chunks.
2280   BasicBlock *DispatchEnter = splitBB(Builder, true);
2281   Value *DispatchCounter;
2282   CanonicalLoopInfo *DispatchCLI = createCanonicalLoop(
2283       {Builder.saveIP(), DL},
2284       [&](InsertPointTy BodyIP, Value *Counter) { DispatchCounter = Counter; },
2285       FirstChunkStart, CastedTripCount, NextChunkStride,
2286       /*IsSigned=*/false, /*InclusiveStop=*/false, /*ComputeIP=*/{},
2287       "dispatch");
2288 
2289   // Remember the BasicBlocks of the dispatch loop we need, then invalidate to
2290   // not have to preserve the canonical invariant.
2291   BasicBlock *DispatchBody = DispatchCLI->getBody();
2292   BasicBlock *DispatchLatch = DispatchCLI->getLatch();
2293   BasicBlock *DispatchExit = DispatchCLI->getExit();
2294   BasicBlock *DispatchAfter = DispatchCLI->getAfter();
2295   DispatchCLI->invalidate();
2296 
2297   // Rewire the original loop to become the chunk loop inside the dispatch loop.
2298   redirectTo(DispatchAfter, CLI->getAfter(), DL);
2299   redirectTo(CLI->getExit(), DispatchLatch, DL);
2300   redirectTo(DispatchBody, DispatchEnter, DL);
2301 
2302   // Prepare the prolog of the chunk loop.
2303   Builder.restoreIP(CLI->getPreheaderIP());
2304   Builder.SetCurrentDebugLocation(DL);
2305 
2306   // Compute the number of iterations of the chunk loop.
2307   Builder.SetInsertPoint(CLI->getPreheader()->getTerminator());
2308   Value *ChunkEnd = Builder.CreateAdd(DispatchCounter, ChunkRange);
2309   Value *IsLastChunk =
2310       Builder.CreateICmpUGE(ChunkEnd, CastedTripCount, "omp_chunk.is_last");
2311   Value *CountUntilOrigTripCount =
2312       Builder.CreateSub(CastedTripCount, DispatchCounter);
2313   Value *ChunkTripCount = Builder.CreateSelect(
2314       IsLastChunk, CountUntilOrigTripCount, ChunkRange, "omp_chunk.tripcount");
2315   Value *BackcastedChunkTC =
2316       Builder.CreateTrunc(ChunkTripCount, IVTy, "omp_chunk.tripcount.trunc");
2317   CLI->setTripCount(BackcastedChunkTC);
2318 
2319   // Update all uses of the induction variable except the one in the condition
2320   // block that compares it with the actual upper bound, and the increment in
2321   // the latch block.
2322   Value *BackcastedDispatchCounter =
2323       Builder.CreateTrunc(DispatchCounter, IVTy, "omp_dispatch.iv.trunc");
2324   CLI->mapIndVar([&](Instruction *) -> Value * {
2325     Builder.restoreIP(CLI->getBodyIP());
2326     return Builder.CreateAdd(IV, BackcastedDispatchCounter);
2327   });
2328 
2329   // In the "exit" block, call the "fini" function.
2330   Builder.SetInsertPoint(DispatchExit, DispatchExit->getFirstInsertionPt());
2331   Builder.CreateCall(StaticFini, {SrcLoc, ThreadNum});
2332 
2333   // Add the barrier if requested.
2334   if (NeedsBarrier)
2335     createBarrier(LocationDescription(Builder.saveIP(), DL), OMPD_for,
2336                   /*ForceSimpleCall=*/false, /*CheckCancelFlag=*/false);
2337 
2338 #ifndef NDEBUG
2339   // Even though we currently do not support applying additional methods to it,
2340   // the chunk loop should remain a canonical loop.
2341   CLI->assertOK();
2342 #endif
2343 
2344   return {DispatchAfter, DispatchAfter->getFirstInsertionPt()};
2345 }
2346 
2347 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyWorkshareLoop(
2348     DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP,
2349     bool NeedsBarrier, llvm::omp::ScheduleKind SchedKind,
2350     llvm::Value *ChunkSize, bool HasSimdModifier, bool HasMonotonicModifier,
2351     bool HasNonmonotonicModifier, bool HasOrderedClause) {
2352   OMPScheduleType EffectiveScheduleType = computeOpenMPScheduleType(
2353       SchedKind, ChunkSize, HasSimdModifier, HasMonotonicModifier,
2354       HasNonmonotonicModifier, HasOrderedClause);
2355 
2356   bool IsOrdered = (EffectiveScheduleType & OMPScheduleType::ModifierOrdered) ==
2357                    OMPScheduleType::ModifierOrdered;
2358   switch (EffectiveScheduleType & ~OMPScheduleType::ModifierMask) {
2359   case OMPScheduleType::BaseStatic:
2360     assert(!ChunkSize && "No chunk size with static-chunked schedule");
2361     if (IsOrdered)
2362       return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, EffectiveScheduleType,
2363                                        NeedsBarrier, ChunkSize);
2364     // FIXME: Monotonicity ignored?
2365     return applyStaticWorkshareLoop(DL, CLI, AllocaIP, NeedsBarrier);
2366 
2367   case OMPScheduleType::BaseStaticChunked:
2368     if (IsOrdered)
2369       return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, EffectiveScheduleType,
2370                                        NeedsBarrier, ChunkSize);
2371     // FIXME: Monotonicity ignored?
2372     return applyStaticChunkedWorkshareLoop(DL, CLI, AllocaIP, NeedsBarrier,
2373                                            ChunkSize);
2374 
2375   case OMPScheduleType::BaseRuntime:
2376   case OMPScheduleType::BaseAuto:
2377   case OMPScheduleType::BaseGreedy:
2378   case OMPScheduleType::BaseBalanced:
2379   case OMPScheduleType::BaseSteal:
2380   case OMPScheduleType::BaseGuidedSimd:
2381   case OMPScheduleType::BaseRuntimeSimd:
2382     assert(!ChunkSize &&
2383            "schedule type does not support user-defined chunk sizes");
2384     LLVM_FALLTHROUGH;
2385   case OMPScheduleType::BaseDynamicChunked:
2386   case OMPScheduleType::BaseGuidedChunked:
2387   case OMPScheduleType::BaseGuidedIterativeChunked:
2388   case OMPScheduleType::BaseGuidedAnalyticalChunked:
2389   case OMPScheduleType::BaseStaticBalancedChunked:
2390     return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, EffectiveScheduleType,
2391                                      NeedsBarrier, ChunkSize);
2392 
2393   default:
2394     llvm_unreachable("Unknown/unimplemented schedule kind");
2395   }
2396 }
2397 
2398 /// Returns an LLVM function to call for initializing loop bounds using OpenMP
2399 /// dynamic scheduling depending on `type`. Only i32 and i64 are supported by
2400 /// the runtime. Always interpret integers as unsigned similarly to
2401 /// CanonicalLoopInfo.
2402 static FunctionCallee
2403 getKmpcForDynamicInitForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) {
2404   unsigned Bitwidth = Ty->getIntegerBitWidth();
2405   if (Bitwidth == 32)
2406     return OMPBuilder.getOrCreateRuntimeFunction(
2407         M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_4u);
2408   if (Bitwidth == 64)
2409     return OMPBuilder.getOrCreateRuntimeFunction(
2410         M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_8u);
2411   llvm_unreachable("unknown OpenMP loop iterator bitwidth");
2412 }
2413 
2414 /// Returns an LLVM function to call for updating the next loop using OpenMP
2415 /// dynamic scheduling depending on `type`. Only i32 and i64 are supported by
2416 /// the runtime. Always interpret integers as unsigned similarly to
2417 /// CanonicalLoopInfo.
2418 static FunctionCallee
2419 getKmpcForDynamicNextForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) {
2420   unsigned Bitwidth = Ty->getIntegerBitWidth();
2421   if (Bitwidth == 32)
2422     return OMPBuilder.getOrCreateRuntimeFunction(
2423         M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_4u);
2424   if (Bitwidth == 64)
2425     return OMPBuilder.getOrCreateRuntimeFunction(
2426         M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_8u);
2427   llvm_unreachable("unknown OpenMP loop iterator bitwidth");
2428 }
2429 
2430 /// Returns an LLVM function to call for finalizing the dynamic loop using
2431 /// depending on `type`. Only i32 and i64 are supported by the runtime. Always
2432 /// interpret integers as unsigned similarly to CanonicalLoopInfo.
2433 static FunctionCallee
2434 getKmpcForDynamicFiniForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) {
2435   unsigned Bitwidth = Ty->getIntegerBitWidth();
2436   if (Bitwidth == 32)
2437     return OMPBuilder.getOrCreateRuntimeFunction(
2438         M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_fini_4u);
2439   if (Bitwidth == 64)
2440     return OMPBuilder.getOrCreateRuntimeFunction(
2441         M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_fini_8u);
2442   llvm_unreachable("unknown OpenMP loop iterator bitwidth");
2443 }
2444 
2445 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyDynamicWorkshareLoop(
2446     DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP,
2447     OMPScheduleType SchedType, bool NeedsBarrier, Value *Chunk) {
2448   assert(CLI->isValid() && "Requires a valid canonical loop");
2449   assert(!isConflictIP(AllocaIP, CLI->getPreheaderIP()) &&
2450          "Require dedicated allocate IP");
2451   assert(isValidWorkshareLoopScheduleType(SchedType) &&
2452          "Require valid schedule type");
2453 
2454   bool Ordered = (SchedType & OMPScheduleType::ModifierOrdered) ==
2455                  OMPScheduleType::ModifierOrdered;
2456 
2457   // Set up the source location value for OpenMP runtime.
2458   Builder.SetCurrentDebugLocation(DL);
2459 
2460   uint32_t SrcLocStrSize;
2461   Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize);
2462   Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
2463 
2464   // Declare useful OpenMP runtime functions.
2465   Value *IV = CLI->getIndVar();
2466   Type *IVTy = IV->getType();
2467   FunctionCallee DynamicInit = getKmpcForDynamicInitForType(IVTy, M, *this);
2468   FunctionCallee DynamicNext = getKmpcForDynamicNextForType(IVTy, M, *this);
2469 
2470   // Allocate space for computed loop bounds as expected by the "init" function.
2471   Builder.restoreIP(AllocaIP);
2472   Type *I32Type = Type::getInt32Ty(M.getContext());
2473   Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter");
2474   Value *PLowerBound = Builder.CreateAlloca(IVTy, nullptr, "p.lowerbound");
2475   Value *PUpperBound = Builder.CreateAlloca(IVTy, nullptr, "p.upperbound");
2476   Value *PStride = Builder.CreateAlloca(IVTy, nullptr, "p.stride");
2477 
2478   // At the end of the preheader, prepare for calling the "init" function by
2479   // storing the current loop bounds into the allocated space. A canonical loop
2480   // always iterates from 0 to trip-count with step 1. Note that "init" expects
2481   // and produces an inclusive upper bound.
2482   BasicBlock *PreHeader = CLI->getPreheader();
2483   Builder.SetInsertPoint(PreHeader->getTerminator());
2484   Constant *One = ConstantInt::get(IVTy, 1);
2485   Builder.CreateStore(One, PLowerBound);
2486   Value *UpperBound = CLI->getTripCount();
2487   Builder.CreateStore(UpperBound, PUpperBound);
2488   Builder.CreateStore(One, PStride);
2489 
2490   BasicBlock *Header = CLI->getHeader();
2491   BasicBlock *Exit = CLI->getExit();
2492   BasicBlock *Cond = CLI->getCond();
2493   BasicBlock *Latch = CLI->getLatch();
2494   InsertPointTy AfterIP = CLI->getAfterIP();
2495 
2496   // The CLI will be "broken" in the code below, as the loop is no longer
2497   // a valid canonical loop.
2498 
2499   if (!Chunk)
2500     Chunk = One;
2501 
2502   Value *ThreadNum = getOrCreateThreadID(SrcLoc);
2503 
2504   Constant *SchedulingType =
2505       ConstantInt::get(I32Type, static_cast<int>(SchedType));
2506 
2507   // Call the "init" function.
2508   Builder.CreateCall(DynamicInit,
2509                      {SrcLoc, ThreadNum, SchedulingType, /* LowerBound */ One,
2510                       UpperBound, /* step */ One, Chunk});
2511 
2512   // An outer loop around the existing one.
2513   BasicBlock *OuterCond = BasicBlock::Create(
2514       PreHeader->getContext(), Twine(PreHeader->getName()) + ".outer.cond",
2515       PreHeader->getParent());
2516   // This needs to be 32-bit always, so can't use the IVTy Zero above.
2517   Builder.SetInsertPoint(OuterCond, OuterCond->getFirstInsertionPt());
2518   Value *Res =
2519       Builder.CreateCall(DynamicNext, {SrcLoc, ThreadNum, PLastIter,
2520                                        PLowerBound, PUpperBound, PStride});
2521   Constant *Zero32 = ConstantInt::get(I32Type, 0);
2522   Value *MoreWork = Builder.CreateCmp(CmpInst::ICMP_NE, Res, Zero32);
2523   Value *LowerBound =
2524       Builder.CreateSub(Builder.CreateLoad(IVTy, PLowerBound), One, "lb");
2525   Builder.CreateCondBr(MoreWork, Header, Exit);
2526 
2527   // Change PHI-node in loop header to use outer cond rather than preheader,
2528   // and set IV to the LowerBound.
2529   Instruction *Phi = &Header->front();
2530   auto *PI = cast<PHINode>(Phi);
2531   PI->setIncomingBlock(0, OuterCond);
2532   PI->setIncomingValue(0, LowerBound);
2533 
2534   // Then set the pre-header to jump to the OuterCond
2535   Instruction *Term = PreHeader->getTerminator();
2536   auto *Br = cast<BranchInst>(Term);
2537   Br->setSuccessor(0, OuterCond);
2538 
2539   // Modify the inner condition:
2540   // * Use the UpperBound returned from the DynamicNext call.
2541   // * jump to the loop outer loop when done with one of the inner loops.
2542   Builder.SetInsertPoint(Cond, Cond->getFirstInsertionPt());
2543   UpperBound = Builder.CreateLoad(IVTy, PUpperBound, "ub");
2544   Instruction *Comp = &*Builder.GetInsertPoint();
2545   auto *CI = cast<CmpInst>(Comp);
2546   CI->setOperand(1, UpperBound);
2547   // Redirect the inner exit to branch to outer condition.
2548   Instruction *Branch = &Cond->back();
2549   auto *BI = cast<BranchInst>(Branch);
2550   assert(BI->getSuccessor(1) == Exit);
2551   BI->setSuccessor(1, OuterCond);
2552 
2553   // Call the "fini" function if "ordered" is present in wsloop directive.
2554   if (Ordered) {
2555     Builder.SetInsertPoint(&Latch->back());
2556     FunctionCallee DynamicFini = getKmpcForDynamicFiniForType(IVTy, M, *this);
2557     Builder.CreateCall(DynamicFini, {SrcLoc, ThreadNum});
2558   }
2559 
2560   // Add the barrier if requested.
2561   if (NeedsBarrier) {
2562     Builder.SetInsertPoint(&Exit->back());
2563     createBarrier(LocationDescription(Builder.saveIP(), DL),
2564                   omp::Directive::OMPD_for, /* ForceSimpleCall */ false,
2565                   /* CheckCancelFlag */ false);
2566   }
2567 
2568   CLI->invalidate();
2569   return AfterIP;
2570 }
2571 
2572 /// Redirect all edges that branch to \p OldTarget to \p NewTarget. That is,
2573 /// after this \p OldTarget will be orphaned.
2574 static void redirectAllPredecessorsTo(BasicBlock *OldTarget,
2575                                       BasicBlock *NewTarget, DebugLoc DL) {
2576   for (BasicBlock *Pred : make_early_inc_range(predecessors(OldTarget)))
2577     redirectTo(Pred, NewTarget, DL);
2578 }
2579 
2580 /// Determine which blocks in \p BBs are reachable from outside and remove the
2581 /// ones that are not reachable from the function.
2582 static void removeUnusedBlocksFromParent(ArrayRef<BasicBlock *> BBs) {
2583   SmallPtrSet<BasicBlock *, 6> BBsToErase{BBs.begin(), BBs.end()};
2584   auto HasRemainingUses = [&BBsToErase](BasicBlock *BB) {
2585     for (Use &U : BB->uses()) {
2586       auto *UseInst = dyn_cast<Instruction>(U.getUser());
2587       if (!UseInst)
2588         continue;
2589       if (BBsToErase.count(UseInst->getParent()))
2590         continue;
2591       return true;
2592     }
2593     return false;
2594   };
2595 
2596   while (true) {
2597     bool Changed = false;
2598     for (BasicBlock *BB : make_early_inc_range(BBsToErase)) {
2599       if (HasRemainingUses(BB)) {
2600         BBsToErase.erase(BB);
2601         Changed = true;
2602       }
2603     }
2604     if (!Changed)
2605       break;
2606   }
2607 
2608   SmallVector<BasicBlock *, 7> BBVec(BBsToErase.begin(), BBsToErase.end());
2609   DeleteDeadBlocks(BBVec);
2610 }
2611 
2612 CanonicalLoopInfo *
2613 OpenMPIRBuilder::collapseLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops,
2614                                InsertPointTy ComputeIP) {
2615   assert(Loops.size() >= 1 && "At least one loop required");
2616   size_t NumLoops = Loops.size();
2617 
2618   // Nothing to do if there is already just one loop.
2619   if (NumLoops == 1)
2620     return Loops.front();
2621 
2622   CanonicalLoopInfo *Outermost = Loops.front();
2623   CanonicalLoopInfo *Innermost = Loops.back();
2624   BasicBlock *OrigPreheader = Outermost->getPreheader();
2625   BasicBlock *OrigAfter = Outermost->getAfter();
2626   Function *F = OrigPreheader->getParent();
2627 
2628   // Loop control blocks that may become orphaned later.
2629   SmallVector<BasicBlock *, 12> OldControlBBs;
2630   OldControlBBs.reserve(6 * Loops.size());
2631   for (CanonicalLoopInfo *Loop : Loops)
2632     Loop->collectControlBlocks(OldControlBBs);
2633 
2634   // Setup the IRBuilder for inserting the trip count computation.
2635   Builder.SetCurrentDebugLocation(DL);
2636   if (ComputeIP.isSet())
2637     Builder.restoreIP(ComputeIP);
2638   else
2639     Builder.restoreIP(Outermost->getPreheaderIP());
2640 
2641   // Derive the collapsed' loop trip count.
2642   // TODO: Find common/largest indvar type.
2643   Value *CollapsedTripCount = nullptr;
2644   for (CanonicalLoopInfo *L : Loops) {
2645     assert(L->isValid() &&
2646            "All loops to collapse must be valid canonical loops");
2647     Value *OrigTripCount = L->getTripCount();
2648     if (!CollapsedTripCount) {
2649       CollapsedTripCount = OrigTripCount;
2650       continue;
2651     }
2652 
2653     // TODO: Enable UndefinedSanitizer to diagnose an overflow here.
2654     CollapsedTripCount = Builder.CreateMul(CollapsedTripCount, OrigTripCount,
2655                                            {}, /*HasNUW=*/true);
2656   }
2657 
2658   // Create the collapsed loop control flow.
2659   CanonicalLoopInfo *Result =
2660       createLoopSkeleton(DL, CollapsedTripCount, F,
2661                          OrigPreheader->getNextNode(), OrigAfter, "collapsed");
2662 
2663   // Build the collapsed loop body code.
2664   // Start with deriving the input loop induction variables from the collapsed
2665   // one, using a divmod scheme. To preserve the original loops' order, the
2666   // innermost loop use the least significant bits.
2667   Builder.restoreIP(Result->getBodyIP());
2668 
2669   Value *Leftover = Result->getIndVar();
2670   SmallVector<Value *> NewIndVars;
2671   NewIndVars.resize(NumLoops);
2672   for (int i = NumLoops - 1; i >= 1; --i) {
2673     Value *OrigTripCount = Loops[i]->getTripCount();
2674 
2675     Value *NewIndVar = Builder.CreateURem(Leftover, OrigTripCount);
2676     NewIndVars[i] = NewIndVar;
2677 
2678     Leftover = Builder.CreateUDiv(Leftover, OrigTripCount);
2679   }
2680   // Outermost loop gets all the remaining bits.
2681   NewIndVars[0] = Leftover;
2682 
2683   // Construct the loop body control flow.
2684   // We progressively construct the branch structure following in direction of
2685   // the control flow, from the leading in-between code, the loop nest body, the
2686   // trailing in-between code, and rejoining the collapsed loop's latch.
2687   // ContinueBlock and ContinuePred keep track of the source(s) of next edge. If
2688   // the ContinueBlock is set, continue with that block. If ContinuePred, use
2689   // its predecessors as sources.
2690   BasicBlock *ContinueBlock = Result->getBody();
2691   BasicBlock *ContinuePred = nullptr;
2692   auto ContinueWith = [&ContinueBlock, &ContinuePred, DL](BasicBlock *Dest,
2693                                                           BasicBlock *NextSrc) {
2694     if (ContinueBlock)
2695       redirectTo(ContinueBlock, Dest, DL);
2696     else
2697       redirectAllPredecessorsTo(ContinuePred, Dest, DL);
2698 
2699     ContinueBlock = nullptr;
2700     ContinuePred = NextSrc;
2701   };
2702 
2703   // The code before the nested loop of each level.
2704   // Because we are sinking it into the nest, it will be executed more often
2705   // that the original loop. More sophisticated schemes could keep track of what
2706   // the in-between code is and instantiate it only once per thread.
2707   for (size_t i = 0; i < NumLoops - 1; ++i)
2708     ContinueWith(Loops[i]->getBody(), Loops[i + 1]->getHeader());
2709 
2710   // Connect the loop nest body.
2711   ContinueWith(Innermost->getBody(), Innermost->getLatch());
2712 
2713   // The code after the nested loop at each level.
2714   for (size_t i = NumLoops - 1; i > 0; --i)
2715     ContinueWith(Loops[i]->getAfter(), Loops[i - 1]->getLatch());
2716 
2717   // Connect the finished loop to the collapsed loop latch.
2718   ContinueWith(Result->getLatch(), nullptr);
2719 
2720   // Replace the input loops with the new collapsed loop.
2721   redirectTo(Outermost->getPreheader(), Result->getPreheader(), DL);
2722   redirectTo(Result->getAfter(), Outermost->getAfter(), DL);
2723 
2724   // Replace the input loop indvars with the derived ones.
2725   for (size_t i = 0; i < NumLoops; ++i)
2726     Loops[i]->getIndVar()->replaceAllUsesWith(NewIndVars[i]);
2727 
2728   // Remove unused parts of the input loops.
2729   removeUnusedBlocksFromParent(OldControlBBs);
2730 
2731   for (CanonicalLoopInfo *L : Loops)
2732     L->invalidate();
2733 
2734 #ifndef NDEBUG
2735   Result->assertOK();
2736 #endif
2737   return Result;
2738 }
2739 
2740 std::vector<CanonicalLoopInfo *>
2741 OpenMPIRBuilder::tileLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops,
2742                            ArrayRef<Value *> TileSizes) {
2743   assert(TileSizes.size() == Loops.size() &&
2744          "Must pass as many tile sizes as there are loops");
2745   int NumLoops = Loops.size();
2746   assert(NumLoops >= 1 && "At least one loop to tile required");
2747 
2748   CanonicalLoopInfo *OutermostLoop = Loops.front();
2749   CanonicalLoopInfo *InnermostLoop = Loops.back();
2750   Function *F = OutermostLoop->getBody()->getParent();
2751   BasicBlock *InnerEnter = InnermostLoop->getBody();
2752   BasicBlock *InnerLatch = InnermostLoop->getLatch();
2753 
2754   // Loop control blocks that may become orphaned later.
2755   SmallVector<BasicBlock *, 12> OldControlBBs;
2756   OldControlBBs.reserve(6 * Loops.size());
2757   for (CanonicalLoopInfo *Loop : Loops)
2758     Loop->collectControlBlocks(OldControlBBs);
2759 
2760   // Collect original trip counts and induction variable to be accessible by
2761   // index. Also, the structure of the original loops is not preserved during
2762   // the construction of the tiled loops, so do it before we scavenge the BBs of
2763   // any original CanonicalLoopInfo.
2764   SmallVector<Value *, 4> OrigTripCounts, OrigIndVars;
2765   for (CanonicalLoopInfo *L : Loops) {
2766     assert(L->isValid() && "All input loops must be valid canonical loops");
2767     OrigTripCounts.push_back(L->getTripCount());
2768     OrigIndVars.push_back(L->getIndVar());
2769   }
2770 
2771   // Collect the code between loop headers. These may contain SSA definitions
2772   // that are used in the loop nest body. To be usable with in the innermost
2773   // body, these BasicBlocks will be sunk into the loop nest body. That is,
2774   // these instructions may be executed more often than before the tiling.
2775   // TODO: It would be sufficient to only sink them into body of the
2776   // corresponding tile loop.
2777   SmallVector<std::pair<BasicBlock *, BasicBlock *>, 4> InbetweenCode;
2778   for (int i = 0; i < NumLoops - 1; ++i) {
2779     CanonicalLoopInfo *Surrounding = Loops[i];
2780     CanonicalLoopInfo *Nested = Loops[i + 1];
2781 
2782     BasicBlock *EnterBB = Surrounding->getBody();
2783     BasicBlock *ExitBB = Nested->getHeader();
2784     InbetweenCode.emplace_back(EnterBB, ExitBB);
2785   }
2786 
2787   // Compute the trip counts of the floor loops.
2788   Builder.SetCurrentDebugLocation(DL);
2789   Builder.restoreIP(OutermostLoop->getPreheaderIP());
2790   SmallVector<Value *, 4> FloorCount, FloorRems;
2791   for (int i = 0; i < NumLoops; ++i) {
2792     Value *TileSize = TileSizes[i];
2793     Value *OrigTripCount = OrigTripCounts[i];
2794     Type *IVType = OrigTripCount->getType();
2795 
2796     Value *FloorTripCount = Builder.CreateUDiv(OrigTripCount, TileSize);
2797     Value *FloorTripRem = Builder.CreateURem(OrigTripCount, TileSize);
2798 
2799     // 0 if tripcount divides the tilesize, 1 otherwise.
2800     // 1 means we need an additional iteration for a partial tile.
2801     //
2802     // Unfortunately we cannot just use the roundup-formula
2803     //   (tripcount + tilesize - 1)/tilesize
2804     // because the summation might overflow. We do not want introduce undefined
2805     // behavior when the untiled loop nest did not.
2806     Value *FloorTripOverflow =
2807         Builder.CreateICmpNE(FloorTripRem, ConstantInt::get(IVType, 0));
2808 
2809     FloorTripOverflow = Builder.CreateZExt(FloorTripOverflow, IVType);
2810     FloorTripCount =
2811         Builder.CreateAdd(FloorTripCount, FloorTripOverflow,
2812                           "omp_floor" + Twine(i) + ".tripcount", true);
2813 
2814     // Remember some values for later use.
2815     FloorCount.push_back(FloorTripCount);
2816     FloorRems.push_back(FloorTripRem);
2817   }
2818 
2819   // Generate the new loop nest, from the outermost to the innermost.
2820   std::vector<CanonicalLoopInfo *> Result;
2821   Result.reserve(NumLoops * 2);
2822 
2823   // The basic block of the surrounding loop that enters the nest generated
2824   // loop.
2825   BasicBlock *Enter = OutermostLoop->getPreheader();
2826 
2827   // The basic block of the surrounding loop where the inner code should
2828   // continue.
2829   BasicBlock *Continue = OutermostLoop->getAfter();
2830 
2831   // Where the next loop basic block should be inserted.
2832   BasicBlock *OutroInsertBefore = InnermostLoop->getExit();
2833 
2834   auto EmbeddNewLoop =
2835       [this, DL, F, InnerEnter, &Enter, &Continue, &OutroInsertBefore](
2836           Value *TripCount, const Twine &Name) -> CanonicalLoopInfo * {
2837     CanonicalLoopInfo *EmbeddedLoop = createLoopSkeleton(
2838         DL, TripCount, F, InnerEnter, OutroInsertBefore, Name);
2839     redirectTo(Enter, EmbeddedLoop->getPreheader(), DL);
2840     redirectTo(EmbeddedLoop->getAfter(), Continue, DL);
2841 
2842     // Setup the position where the next embedded loop connects to this loop.
2843     Enter = EmbeddedLoop->getBody();
2844     Continue = EmbeddedLoop->getLatch();
2845     OutroInsertBefore = EmbeddedLoop->getLatch();
2846     return EmbeddedLoop;
2847   };
2848 
2849   auto EmbeddNewLoops = [&Result, &EmbeddNewLoop](ArrayRef<Value *> TripCounts,
2850                                                   const Twine &NameBase) {
2851     for (auto P : enumerate(TripCounts)) {
2852       CanonicalLoopInfo *EmbeddedLoop =
2853           EmbeddNewLoop(P.value(), NameBase + Twine(P.index()));
2854       Result.push_back(EmbeddedLoop);
2855     }
2856   };
2857 
2858   EmbeddNewLoops(FloorCount, "floor");
2859 
2860   // Within the innermost floor loop, emit the code that computes the tile
2861   // sizes.
2862   Builder.SetInsertPoint(Enter->getTerminator());
2863   SmallVector<Value *, 4> TileCounts;
2864   for (int i = 0; i < NumLoops; ++i) {
2865     CanonicalLoopInfo *FloorLoop = Result[i];
2866     Value *TileSize = TileSizes[i];
2867 
2868     Value *FloorIsEpilogue =
2869         Builder.CreateICmpEQ(FloorLoop->getIndVar(), FloorCount[i]);
2870     Value *TileTripCount =
2871         Builder.CreateSelect(FloorIsEpilogue, FloorRems[i], TileSize);
2872 
2873     TileCounts.push_back(TileTripCount);
2874   }
2875 
2876   // Create the tile loops.
2877   EmbeddNewLoops(TileCounts, "tile");
2878 
2879   // Insert the inbetween code into the body.
2880   BasicBlock *BodyEnter = Enter;
2881   BasicBlock *BodyEntered = nullptr;
2882   for (std::pair<BasicBlock *, BasicBlock *> P : InbetweenCode) {
2883     BasicBlock *EnterBB = P.first;
2884     BasicBlock *ExitBB = P.second;
2885 
2886     if (BodyEnter)
2887       redirectTo(BodyEnter, EnterBB, DL);
2888     else
2889       redirectAllPredecessorsTo(BodyEntered, EnterBB, DL);
2890 
2891     BodyEnter = nullptr;
2892     BodyEntered = ExitBB;
2893   }
2894 
2895   // Append the original loop nest body into the generated loop nest body.
2896   if (BodyEnter)
2897     redirectTo(BodyEnter, InnerEnter, DL);
2898   else
2899     redirectAllPredecessorsTo(BodyEntered, InnerEnter, DL);
2900   redirectAllPredecessorsTo(InnerLatch, Continue, DL);
2901 
2902   // Replace the original induction variable with an induction variable computed
2903   // from the tile and floor induction variables.
2904   Builder.restoreIP(Result.back()->getBodyIP());
2905   for (int i = 0; i < NumLoops; ++i) {
2906     CanonicalLoopInfo *FloorLoop = Result[i];
2907     CanonicalLoopInfo *TileLoop = Result[NumLoops + i];
2908     Value *OrigIndVar = OrigIndVars[i];
2909     Value *Size = TileSizes[i];
2910 
2911     Value *Scale =
2912         Builder.CreateMul(Size, FloorLoop->getIndVar(), {}, /*HasNUW=*/true);
2913     Value *Shift =
2914         Builder.CreateAdd(Scale, TileLoop->getIndVar(), {}, /*HasNUW=*/true);
2915     OrigIndVar->replaceAllUsesWith(Shift);
2916   }
2917 
2918   // Remove unused parts of the original loops.
2919   removeUnusedBlocksFromParent(OldControlBBs);
2920 
2921   for (CanonicalLoopInfo *L : Loops)
2922     L->invalidate();
2923 
2924 #ifndef NDEBUG
2925   for (CanonicalLoopInfo *GenL : Result)
2926     GenL->assertOK();
2927 #endif
2928   return Result;
2929 }
2930 
2931 /// Attach metadata \p Properties to the basic block described by \p BB. If the
2932 /// basic block already has metadata, the basic block properties are appended.
2933 static void addBasicBlockMetadata(BasicBlock *BB,
2934                                   ArrayRef<Metadata *> Properties) {
2935   // Nothing to do if no property to attach.
2936   if (Properties.empty())
2937     return;
2938 
2939   LLVMContext &Ctx = BB->getContext();
2940   SmallVector<Metadata *> NewProperties;
2941   NewProperties.push_back(nullptr);
2942 
2943   // If the basic block already has metadata, prepend it to the new metadata.
2944   MDNode *Existing = BB->getTerminator()->getMetadata(LLVMContext::MD_loop);
2945   if (Existing)
2946     append_range(NewProperties, drop_begin(Existing->operands(), 1));
2947 
2948   append_range(NewProperties, Properties);
2949   MDNode *BasicBlockID = MDNode::getDistinct(Ctx, NewProperties);
2950   BasicBlockID->replaceOperandWith(0, BasicBlockID);
2951 
2952   BB->getTerminator()->setMetadata(LLVMContext::MD_loop, BasicBlockID);
2953 }
2954 
2955 /// Attach loop metadata \p Properties to the loop described by \p Loop. If the
2956 /// loop already has metadata, the loop properties are appended.
2957 static void addLoopMetadata(CanonicalLoopInfo *Loop,
2958                             ArrayRef<Metadata *> Properties) {
2959   assert(Loop->isValid() && "Expecting a valid CanonicalLoopInfo");
2960 
2961   // Attach metadata to the loop's latch
2962   BasicBlock *Latch = Loop->getLatch();
2963   assert(Latch && "A valid CanonicalLoopInfo must have a unique latch");
2964   addBasicBlockMetadata(Latch, Properties);
2965 }
2966 
2967 /// Attach llvm.access.group metadata to the memref instructions of \p Block
2968 static void addSimdMetadata(BasicBlock *Block, MDNode *AccessGroup,
2969                             LoopInfo &LI) {
2970   for (Instruction &I : *Block) {
2971     if (I.mayReadOrWriteMemory()) {
2972       // TODO: This instruction may already have access group from
2973       // other pragmas e.g. #pragma clang loop vectorize.  Append
2974       // so that the existing metadata is not overwritten.
2975       I.setMetadata(LLVMContext::MD_access_group, AccessGroup);
2976     }
2977   }
2978 }
2979 
2980 void OpenMPIRBuilder::unrollLoopFull(DebugLoc, CanonicalLoopInfo *Loop) {
2981   LLVMContext &Ctx = Builder.getContext();
2982   addLoopMetadata(
2983       Loop, {MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")),
2984              MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.full"))});
2985 }
2986 
2987 void OpenMPIRBuilder::unrollLoopHeuristic(DebugLoc, CanonicalLoopInfo *Loop) {
2988   LLVMContext &Ctx = Builder.getContext();
2989   addLoopMetadata(
2990       Loop, {
2991                 MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")),
2992             });
2993 }
2994 
2995 void OpenMPIRBuilder::createIfVersion(CanonicalLoopInfo *CanonicalLoop,
2996                                       Value *IfCond, ValueToValueMapTy &VMap,
2997                                       const Twine &NamePrefix) {
2998   Function *F = CanonicalLoop->getFunction();
2999 
3000   // Define where if branch should be inserted
3001   Instruction *SplitBefore;
3002   if (Instruction::classof(IfCond)) {
3003     SplitBefore = dyn_cast<Instruction>(IfCond);
3004   } else {
3005     SplitBefore = CanonicalLoop->getPreheader()->getTerminator();
3006   }
3007 
3008   // TODO: We should not rely on pass manager. Currently we use pass manager
3009   // only for getting llvm::Loop which corresponds to given CanonicalLoopInfo
3010   // object. We should have a method  which returns all blocks between
3011   // CanonicalLoopInfo::getHeader() and CanonicalLoopInfo::getAfter()
3012   FunctionAnalysisManager FAM;
3013   FAM.registerPass([]() { return DominatorTreeAnalysis(); });
3014   FAM.registerPass([]() { return LoopAnalysis(); });
3015   FAM.registerPass([]() { return PassInstrumentationAnalysis(); });
3016 
3017   // Get the loop which needs to be cloned
3018   LoopAnalysis LIA;
3019   LoopInfo &&LI = LIA.run(*F, FAM);
3020   Loop *L = LI.getLoopFor(CanonicalLoop->getHeader());
3021 
3022   // Create additional blocks for the if statement
3023   BasicBlock *Head = SplitBefore->getParent();
3024   Instruction *HeadOldTerm = Head->getTerminator();
3025   llvm::LLVMContext &C = Head->getContext();
3026   llvm::BasicBlock *ThenBlock = llvm::BasicBlock::Create(
3027       C, NamePrefix + ".if.then", Head->getParent(), Head->getNextNode());
3028   llvm::BasicBlock *ElseBlock = llvm::BasicBlock::Create(
3029       C, NamePrefix + ".if.else", Head->getParent(), CanonicalLoop->getExit());
3030 
3031   // Create if condition branch.
3032   Builder.SetInsertPoint(HeadOldTerm);
3033   Instruction *BrInstr =
3034       Builder.CreateCondBr(IfCond, ThenBlock, /*ifFalse*/ ElseBlock);
3035   InsertPointTy IP{BrInstr->getParent(), ++BrInstr->getIterator()};
3036   // Then block contains branch to omp loop which needs to be vectorized
3037   spliceBB(IP, ThenBlock, false);
3038   ThenBlock->replaceSuccessorsPhiUsesWith(Head, ThenBlock);
3039 
3040   Builder.SetInsertPoint(ElseBlock);
3041 
3042   // Clone loop for the else branch
3043   SmallVector<BasicBlock *, 8> NewBlocks;
3044 
3045   VMap[CanonicalLoop->getPreheader()] = ElseBlock;
3046   for (BasicBlock *Block : L->getBlocks()) {
3047     BasicBlock *NewBB = CloneBasicBlock(Block, VMap, "", F);
3048     NewBB->moveBefore(CanonicalLoop->getExit());
3049     VMap[Block] = NewBB;
3050     NewBlocks.push_back(NewBB);
3051   }
3052   remapInstructionsInBlocks(NewBlocks, VMap);
3053   Builder.CreateBr(NewBlocks.front());
3054 }
3055 
3056 void OpenMPIRBuilder::applySimd(CanonicalLoopInfo *CanonicalLoop,
3057                                 MapVector<Value *, Value *> AlignedVars,
3058                                 Value *IfCond, OrderKind Order,
3059                                 ConstantInt *Simdlen, ConstantInt *Safelen) {
3060   LLVMContext &Ctx = Builder.getContext();
3061 
3062   Function *F = CanonicalLoop->getFunction();
3063 
3064   // TODO: We should not rely on pass manager. Currently we use pass manager
3065   // only for getting llvm::Loop which corresponds to given CanonicalLoopInfo
3066   // object. We should have a method  which returns all blocks between
3067   // CanonicalLoopInfo::getHeader() and CanonicalLoopInfo::getAfter()
3068   FunctionAnalysisManager FAM;
3069   FAM.registerPass([]() { return DominatorTreeAnalysis(); });
3070   FAM.registerPass([]() { return LoopAnalysis(); });
3071   FAM.registerPass([]() { return PassInstrumentationAnalysis(); });
3072 
3073   LoopAnalysis LIA;
3074   LoopInfo &&LI = LIA.run(*F, FAM);
3075 
3076   Loop *L = LI.getLoopFor(CanonicalLoop->getHeader());
3077   if (AlignedVars.size()) {
3078     InsertPointTy IP = Builder.saveIP();
3079     Builder.SetInsertPoint(CanonicalLoop->getPreheader()->getTerminator());
3080     for (auto &AlignedItem : AlignedVars) {
3081       Value *AlignedPtr = AlignedItem.first;
3082       Value *Alignment = AlignedItem.second;
3083       Builder.CreateAlignmentAssumption(F->getParent()->getDataLayout(),
3084                                         AlignedPtr, Alignment);
3085     }
3086     Builder.restoreIP(IP);
3087   }
3088 
3089   if (IfCond) {
3090     ValueToValueMapTy VMap;
3091     createIfVersion(CanonicalLoop, IfCond, VMap, "simd");
3092     // Add metadata to the cloned loop which disables vectorization
3093     Value *MappedLatch = VMap.lookup(CanonicalLoop->getLatch());
3094     assert(MappedLatch &&
3095            "Cannot find value which corresponds to original loop latch");
3096     assert(isa<BasicBlock>(MappedLatch) &&
3097            "Cannot cast mapped latch block value to BasicBlock");
3098     BasicBlock *NewLatchBlock = dyn_cast<BasicBlock>(MappedLatch);
3099     ConstantAsMetadata *BoolConst =
3100         ConstantAsMetadata::get(ConstantInt::getFalse(Type::getInt1Ty(Ctx)));
3101     addBasicBlockMetadata(
3102         NewLatchBlock,
3103         {MDNode::get(Ctx, {MDString::get(Ctx, "llvm.loop.vectorize.enable"),
3104                            BoolConst})});
3105   }
3106 
3107   SmallSet<BasicBlock *, 8> Reachable;
3108 
3109   // Get the basic blocks from the loop in which memref instructions
3110   // can be found.
3111   // TODO: Generalize getting all blocks inside a CanonicalizeLoopInfo,
3112   // preferably without running any passes.
3113   for (BasicBlock *Block : L->getBlocks()) {
3114     if (Block == CanonicalLoop->getCond() ||
3115         Block == CanonicalLoop->getHeader())
3116       continue;
3117     Reachable.insert(Block);
3118   }
3119 
3120   SmallVector<Metadata *> LoopMDList;
3121 
3122   // In presence of finite 'safelen', it may be unsafe to mark all
3123   // the memory instructions parallel, because loop-carried
3124   // dependences of 'safelen' iterations are possible.
3125   // If clause order(concurrent) is specified then the memory instructions
3126   // are marked parallel even if 'safelen' is finite.
3127   if ((Safelen == nullptr) || (Order == OrderKind::OMP_ORDER_concurrent)) {
3128     // Add access group metadata to memory-access instructions.
3129     MDNode *AccessGroup = MDNode::getDistinct(Ctx, {});
3130     for (BasicBlock *BB : Reachable)
3131       addSimdMetadata(BB, AccessGroup, LI);
3132     // TODO:  If the loop has existing parallel access metadata, have
3133     // to combine two lists.
3134     LoopMDList.push_back(MDNode::get(
3135         Ctx, {MDString::get(Ctx, "llvm.loop.parallel_accesses"), AccessGroup}));
3136   }
3137 
3138   // Use the above access group metadata to create loop level
3139   // metadata, which should be distinct for each loop.
3140   ConstantAsMetadata *BoolConst =
3141       ConstantAsMetadata::get(ConstantInt::getTrue(Type::getInt1Ty(Ctx)));
3142   LoopMDList.push_back(MDNode::get(
3143       Ctx, {MDString::get(Ctx, "llvm.loop.vectorize.enable"), BoolConst}));
3144 
3145   if (Simdlen || Safelen) {
3146     // If both simdlen and safelen clauses are specified, the value of the
3147     // simdlen parameter must be less than or equal to the value of the safelen
3148     // parameter. Therefore, use safelen only in the absence of simdlen.
3149     ConstantInt *VectorizeWidth = Simdlen == nullptr ? Safelen : Simdlen;
3150     LoopMDList.push_back(
3151         MDNode::get(Ctx, {MDString::get(Ctx, "llvm.loop.vectorize.width"),
3152                           ConstantAsMetadata::get(VectorizeWidth)}));
3153   }
3154 
3155   addLoopMetadata(CanonicalLoop, LoopMDList);
3156 }
3157 
3158 /// Create the TargetMachine object to query the backend for optimization
3159 /// preferences.
3160 ///
3161 /// Ideally, this would be passed from the front-end to the OpenMPBuilder, but
3162 /// e.g. Clang does not pass it to its CodeGen layer and creates it only when
3163 /// needed for the LLVM pass pipline. We use some default options to avoid
3164 /// having to pass too many settings from the frontend that probably do not
3165 /// matter.
3166 ///
3167 /// Currently, TargetMachine is only used sometimes by the unrollLoopPartial
3168 /// method. If we are going to use TargetMachine for more purposes, especially
3169 /// those that are sensitive to TargetOptions, RelocModel and CodeModel, it
3170 /// might become be worth requiring front-ends to pass on their TargetMachine,
3171 /// or at least cache it between methods. Note that while fontends such as Clang
3172 /// have just a single main TargetMachine per translation unit, "target-cpu" and
3173 /// "target-features" that determine the TargetMachine are per-function and can
3174 /// be overrided using __attribute__((target("OPTIONS"))).
3175 static std::unique_ptr<TargetMachine>
3176 createTargetMachine(Function *F, CodeGenOpt::Level OptLevel) {
3177   Module *M = F->getParent();
3178 
3179   StringRef CPU = F->getFnAttribute("target-cpu").getValueAsString();
3180   StringRef Features = F->getFnAttribute("target-features").getValueAsString();
3181   const std::string &Triple = M->getTargetTriple();
3182 
3183   std::string Error;
3184   const llvm::Target *TheTarget = TargetRegistry::lookupTarget(Triple, Error);
3185   if (!TheTarget)
3186     return {};
3187 
3188   llvm::TargetOptions Options;
3189   return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine(
3190       Triple, CPU, Features, Options, /*RelocModel=*/std::nullopt,
3191       /*CodeModel=*/std::nullopt, OptLevel));
3192 }
3193 
3194 /// Heuristically determine the best-performant unroll factor for \p CLI. This
3195 /// depends on the target processor. We are re-using the same heuristics as the
3196 /// LoopUnrollPass.
3197 static int32_t computeHeuristicUnrollFactor(CanonicalLoopInfo *CLI) {
3198   Function *F = CLI->getFunction();
3199 
3200   // Assume the user requests the most aggressive unrolling, even if the rest of
3201   // the code is optimized using a lower setting.
3202   CodeGenOpt::Level OptLevel = CodeGenOpt::Aggressive;
3203   std::unique_ptr<TargetMachine> TM = createTargetMachine(F, OptLevel);
3204 
3205   FunctionAnalysisManager FAM;
3206   FAM.registerPass([]() { return TargetLibraryAnalysis(); });
3207   FAM.registerPass([]() { return AssumptionAnalysis(); });
3208   FAM.registerPass([]() { return DominatorTreeAnalysis(); });
3209   FAM.registerPass([]() { return LoopAnalysis(); });
3210   FAM.registerPass([]() { return ScalarEvolutionAnalysis(); });
3211   FAM.registerPass([]() { return PassInstrumentationAnalysis(); });
3212   TargetIRAnalysis TIRA;
3213   if (TM)
3214     TIRA = TargetIRAnalysis(
3215         [&](const Function &F) { return TM->getTargetTransformInfo(F); });
3216   FAM.registerPass([&]() { return TIRA; });
3217 
3218   TargetIRAnalysis::Result &&TTI = TIRA.run(*F, FAM);
3219   ScalarEvolutionAnalysis SEA;
3220   ScalarEvolution &&SE = SEA.run(*F, FAM);
3221   DominatorTreeAnalysis DTA;
3222   DominatorTree &&DT = DTA.run(*F, FAM);
3223   LoopAnalysis LIA;
3224   LoopInfo &&LI = LIA.run(*F, FAM);
3225   AssumptionAnalysis ACT;
3226   AssumptionCache &&AC = ACT.run(*F, FAM);
3227   OptimizationRemarkEmitter ORE{F};
3228 
3229   Loop *L = LI.getLoopFor(CLI->getHeader());
3230   assert(L && "Expecting CanonicalLoopInfo to be recognized as a loop");
3231 
3232   TargetTransformInfo::UnrollingPreferences UP =
3233       gatherUnrollingPreferences(L, SE, TTI,
3234                                  /*BlockFrequencyInfo=*/nullptr,
3235                                  /*ProfileSummaryInfo=*/nullptr, ORE, OptLevel,
3236                                  /*UserThreshold=*/std::nullopt,
3237                                  /*UserCount=*/std::nullopt,
3238                                  /*UserAllowPartial=*/true,
3239                                  /*UserAllowRuntime=*/true,
3240                                  /*UserUpperBound=*/std::nullopt,
3241                                  /*UserFullUnrollMaxCount=*/std::nullopt);
3242 
3243   UP.Force = true;
3244 
3245   // Account for additional optimizations taking place before the LoopUnrollPass
3246   // would unroll the loop.
3247   UP.Threshold *= UnrollThresholdFactor;
3248   UP.PartialThreshold *= UnrollThresholdFactor;
3249 
3250   // Use normal unroll factors even if the rest of the code is optimized for
3251   // size.
3252   UP.OptSizeThreshold = UP.Threshold;
3253   UP.PartialOptSizeThreshold = UP.PartialThreshold;
3254 
3255   LLVM_DEBUG(dbgs() << "Unroll heuristic thresholds:\n"
3256                     << "  Threshold=" << UP.Threshold << "\n"
3257                     << "  PartialThreshold=" << UP.PartialThreshold << "\n"
3258                     << "  OptSizeThreshold=" << UP.OptSizeThreshold << "\n"
3259                     << "  PartialOptSizeThreshold="
3260                     << UP.PartialOptSizeThreshold << "\n");
3261 
3262   // Disable peeling.
3263   TargetTransformInfo::PeelingPreferences PP =
3264       gatherPeelingPreferences(L, SE, TTI,
3265                                /*UserAllowPeeling=*/false,
3266                                /*UserAllowProfileBasedPeeling=*/false,
3267                                /*UnrollingSpecficValues=*/false);
3268 
3269   SmallPtrSet<const Value *, 32> EphValues;
3270   CodeMetrics::collectEphemeralValues(L, &AC, EphValues);
3271 
3272   // Assume that reads and writes to stack variables can be eliminated by
3273   // Mem2Reg, SROA or LICM. That is, don't count them towards the loop body's
3274   // size.
3275   for (BasicBlock *BB : L->blocks()) {
3276     for (Instruction &I : *BB) {
3277       Value *Ptr;
3278       if (auto *Load = dyn_cast<LoadInst>(&I)) {
3279         Ptr = Load->getPointerOperand();
3280       } else if (auto *Store = dyn_cast<StoreInst>(&I)) {
3281         Ptr = Store->getPointerOperand();
3282       } else
3283         continue;
3284 
3285       Ptr = Ptr->stripPointerCasts();
3286 
3287       if (auto *Alloca = dyn_cast<AllocaInst>(Ptr)) {
3288         if (Alloca->getParent() == &F->getEntryBlock())
3289           EphValues.insert(&I);
3290       }
3291     }
3292   }
3293 
3294   unsigned NumInlineCandidates;
3295   bool NotDuplicatable;
3296   bool Convergent;
3297   InstructionCost LoopSizeIC =
3298       ApproximateLoopSize(L, NumInlineCandidates, NotDuplicatable, Convergent,
3299                           TTI, EphValues, UP.BEInsns);
3300   LLVM_DEBUG(dbgs() << "Estimated loop size is " << LoopSizeIC << "\n");
3301 
3302   // Loop is not unrollable if the loop contains certain instructions.
3303   if (NotDuplicatable || Convergent || !LoopSizeIC.isValid()) {
3304     LLVM_DEBUG(dbgs() << "Loop not considered unrollable\n");
3305     return 1;
3306   }
3307   unsigned LoopSize = *LoopSizeIC.getValue();
3308 
3309   // TODO: Determine trip count of \p CLI if constant, computeUnrollCount might
3310   // be able to use it.
3311   int TripCount = 0;
3312   int MaxTripCount = 0;
3313   bool MaxOrZero = false;
3314   unsigned TripMultiple = 0;
3315 
3316   bool UseUpperBound = false;
3317   computeUnrollCount(L, TTI, DT, &LI, &AC, SE, EphValues, &ORE, TripCount,
3318                      MaxTripCount, MaxOrZero, TripMultiple, LoopSize, UP, PP,
3319                      UseUpperBound);
3320   unsigned Factor = UP.Count;
3321   LLVM_DEBUG(dbgs() << "Suggesting unroll factor of " << Factor << "\n");
3322 
3323   // This function returns 1 to signal to not unroll a loop.
3324   if (Factor == 0)
3325     return 1;
3326   return Factor;
3327 }
3328 
3329 void OpenMPIRBuilder::unrollLoopPartial(DebugLoc DL, CanonicalLoopInfo *Loop,
3330                                         int32_t Factor,
3331                                         CanonicalLoopInfo **UnrolledCLI) {
3332   assert(Factor >= 0 && "Unroll factor must not be negative");
3333 
3334   Function *F = Loop->getFunction();
3335   LLVMContext &Ctx = F->getContext();
3336 
3337   // If the unrolled loop is not used for another loop-associated directive, it
3338   // is sufficient to add metadata for the LoopUnrollPass.
3339   if (!UnrolledCLI) {
3340     SmallVector<Metadata *, 2> LoopMetadata;
3341     LoopMetadata.push_back(
3342         MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")));
3343 
3344     if (Factor >= 1) {
3345       ConstantAsMetadata *FactorConst = ConstantAsMetadata::get(
3346           ConstantInt::get(Type::getInt32Ty(Ctx), APInt(32, Factor)));
3347       LoopMetadata.push_back(MDNode::get(
3348           Ctx, {MDString::get(Ctx, "llvm.loop.unroll.count"), FactorConst}));
3349     }
3350 
3351     addLoopMetadata(Loop, LoopMetadata);
3352     return;
3353   }
3354 
3355   // Heuristically determine the unroll factor.
3356   if (Factor == 0)
3357     Factor = computeHeuristicUnrollFactor(Loop);
3358 
3359   // No change required with unroll factor 1.
3360   if (Factor == 1) {
3361     *UnrolledCLI = Loop;
3362     return;
3363   }
3364 
3365   assert(Factor >= 2 &&
3366          "unrolling only makes sense with a factor of 2 or larger");
3367 
3368   Type *IndVarTy = Loop->getIndVarType();
3369 
3370   // Apply partial unrolling by tiling the loop by the unroll-factor, then fully
3371   // unroll the inner loop.
3372   Value *FactorVal =
3373       ConstantInt::get(IndVarTy, APInt(IndVarTy->getIntegerBitWidth(), Factor,
3374                                        /*isSigned=*/false));
3375   std::vector<CanonicalLoopInfo *> LoopNest =
3376       tileLoops(DL, {Loop}, {FactorVal});
3377   assert(LoopNest.size() == 2 && "Expect 2 loops after tiling");
3378   *UnrolledCLI = LoopNest[0];
3379   CanonicalLoopInfo *InnerLoop = LoopNest[1];
3380 
3381   // LoopUnrollPass can only fully unroll loops with constant trip count.
3382   // Unroll by the unroll factor with a fallback epilog for the remainder
3383   // iterations if necessary.
3384   ConstantAsMetadata *FactorConst = ConstantAsMetadata::get(
3385       ConstantInt::get(Type::getInt32Ty(Ctx), APInt(32, Factor)));
3386   addLoopMetadata(
3387       InnerLoop,
3388       {MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")),
3389        MDNode::get(
3390            Ctx, {MDString::get(Ctx, "llvm.loop.unroll.count"), FactorConst})});
3391 
3392 #ifndef NDEBUG
3393   (*UnrolledCLI)->assertOK();
3394 #endif
3395 }
3396 
3397 OpenMPIRBuilder::InsertPointTy
3398 OpenMPIRBuilder::createCopyPrivate(const LocationDescription &Loc,
3399                                    llvm::Value *BufSize, llvm::Value *CpyBuf,
3400                                    llvm::Value *CpyFn, llvm::Value *DidIt) {
3401   if (!updateToLocation(Loc))
3402     return Loc.IP;
3403 
3404   uint32_t SrcLocStrSize;
3405   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3406   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3407   Value *ThreadId = getOrCreateThreadID(Ident);
3408 
3409   llvm::Value *DidItLD = Builder.CreateLoad(Builder.getInt32Ty(), DidIt);
3410 
3411   Value *Args[] = {Ident, ThreadId, BufSize, CpyBuf, CpyFn, DidItLD};
3412 
3413   Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_copyprivate);
3414   Builder.CreateCall(Fn, Args);
3415 
3416   return Builder.saveIP();
3417 }
3418 
3419 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createSingle(
3420     const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB,
3421     FinalizeCallbackTy FiniCB, bool IsNowait, llvm::Value *DidIt) {
3422 
3423   if (!updateToLocation(Loc))
3424     return Loc.IP;
3425 
3426   // If needed (i.e. not null), initialize `DidIt` with 0
3427   if (DidIt) {
3428     Builder.CreateStore(Builder.getInt32(0), DidIt);
3429   }
3430 
3431   Directive OMPD = Directive::OMPD_single;
3432   uint32_t SrcLocStrSize;
3433   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3434   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3435   Value *ThreadId = getOrCreateThreadID(Ident);
3436   Value *Args[] = {Ident, ThreadId};
3437 
3438   Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_single);
3439   Instruction *EntryCall = Builder.CreateCall(EntryRTLFn, Args);
3440 
3441   Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_single);
3442   Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, Args);
3443 
3444   // generates the following:
3445   // if (__kmpc_single()) {
3446   //		.... single region ...
3447   // 		__kmpc_end_single
3448   // }
3449   // __kmpc_barrier
3450 
3451   EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
3452                        /*Conditional*/ true,
3453                        /*hasFinalize*/ true);
3454   if (!IsNowait)
3455     createBarrier(LocationDescription(Builder.saveIP(), Loc.DL),
3456                   omp::Directive::OMPD_unknown, /* ForceSimpleCall */ false,
3457                   /* CheckCancelFlag */ false);
3458   return Builder.saveIP();
3459 }
3460 
3461 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createCritical(
3462     const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB,
3463     FinalizeCallbackTy FiniCB, StringRef CriticalName, Value *HintInst) {
3464 
3465   if (!updateToLocation(Loc))
3466     return Loc.IP;
3467 
3468   Directive OMPD = Directive::OMPD_critical;
3469   uint32_t SrcLocStrSize;
3470   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3471   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3472   Value *ThreadId = getOrCreateThreadID(Ident);
3473   Value *LockVar = getOMPCriticalRegionLock(CriticalName);
3474   Value *Args[] = {Ident, ThreadId, LockVar};
3475 
3476   SmallVector<llvm::Value *, 4> EnterArgs(std::begin(Args), std::end(Args));
3477   Function *RTFn = nullptr;
3478   if (HintInst) {
3479     // Add Hint to entry Args and create call
3480     EnterArgs.push_back(HintInst);
3481     RTFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_critical_with_hint);
3482   } else {
3483     RTFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_critical);
3484   }
3485   Instruction *EntryCall = Builder.CreateCall(RTFn, EnterArgs);
3486 
3487   Function *ExitRTLFn =
3488       getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_critical);
3489   Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, Args);
3490 
3491   return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
3492                               /*Conditional*/ false, /*hasFinalize*/ true);
3493 }
3494 
3495 OpenMPIRBuilder::InsertPointTy
3496 OpenMPIRBuilder::createOrderedDepend(const LocationDescription &Loc,
3497                                      InsertPointTy AllocaIP, unsigned NumLoops,
3498                                      ArrayRef<llvm::Value *> StoreValues,
3499                                      const Twine &Name, bool IsDependSource) {
3500   assert(
3501       llvm::all_of(StoreValues,
3502                    [](Value *SV) { return SV->getType()->isIntegerTy(64); }) &&
3503       "OpenMP runtime requires depend vec with i64 type");
3504 
3505   if (!updateToLocation(Loc))
3506     return Loc.IP;
3507 
3508   // Allocate space for vector and generate alloc instruction.
3509   auto *ArrI64Ty = ArrayType::get(Int64, NumLoops);
3510   Builder.restoreIP(AllocaIP);
3511   AllocaInst *ArgsBase = Builder.CreateAlloca(ArrI64Ty, nullptr, Name);
3512   ArgsBase->setAlignment(Align(8));
3513   Builder.restoreIP(Loc.IP);
3514 
3515   // Store the index value with offset in depend vector.
3516   for (unsigned I = 0; I < NumLoops; ++I) {
3517     Value *DependAddrGEPIter = Builder.CreateInBoundsGEP(
3518         ArrI64Ty, ArgsBase, {Builder.getInt64(0), Builder.getInt64(I)});
3519     StoreInst *STInst = Builder.CreateStore(StoreValues[I], DependAddrGEPIter);
3520     STInst->setAlignment(Align(8));
3521   }
3522 
3523   Value *DependBaseAddrGEP = Builder.CreateInBoundsGEP(
3524       ArrI64Ty, ArgsBase, {Builder.getInt64(0), Builder.getInt64(0)});
3525 
3526   uint32_t SrcLocStrSize;
3527   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3528   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3529   Value *ThreadId = getOrCreateThreadID(Ident);
3530   Value *Args[] = {Ident, ThreadId, DependBaseAddrGEP};
3531 
3532   Function *RTLFn = nullptr;
3533   if (IsDependSource)
3534     RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_doacross_post);
3535   else
3536     RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_doacross_wait);
3537   Builder.CreateCall(RTLFn, Args);
3538 
3539   return Builder.saveIP();
3540 }
3541 
3542 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createOrderedThreadsSimd(
3543     const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB,
3544     FinalizeCallbackTy FiniCB, bool IsThreads) {
3545   if (!updateToLocation(Loc))
3546     return Loc.IP;
3547 
3548   Directive OMPD = Directive::OMPD_ordered;
3549   Instruction *EntryCall = nullptr;
3550   Instruction *ExitCall = nullptr;
3551 
3552   if (IsThreads) {
3553     uint32_t SrcLocStrSize;
3554     Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3555     Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3556     Value *ThreadId = getOrCreateThreadID(Ident);
3557     Value *Args[] = {Ident, ThreadId};
3558 
3559     Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_ordered);
3560     EntryCall = Builder.CreateCall(EntryRTLFn, Args);
3561 
3562     Function *ExitRTLFn =
3563         getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_ordered);
3564     ExitCall = Builder.CreateCall(ExitRTLFn, Args);
3565   }
3566 
3567   return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB,
3568                               /*Conditional*/ false, /*hasFinalize*/ true);
3569 }
3570 
3571 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::EmitOMPInlinedRegion(
3572     Directive OMPD, Instruction *EntryCall, Instruction *ExitCall,
3573     BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB, bool Conditional,
3574     bool HasFinalize, bool IsCancellable) {
3575 
3576   if (HasFinalize)
3577     FinalizationStack.push_back({FiniCB, OMPD, IsCancellable});
3578 
3579   // Create inlined region's entry and body blocks, in preparation
3580   // for conditional creation
3581   BasicBlock *EntryBB = Builder.GetInsertBlock();
3582   Instruction *SplitPos = EntryBB->getTerminator();
3583   if (!isa_and_nonnull<BranchInst>(SplitPos))
3584     SplitPos = new UnreachableInst(Builder.getContext(), EntryBB);
3585   BasicBlock *ExitBB = EntryBB->splitBasicBlock(SplitPos, "omp_region.end");
3586   BasicBlock *FiniBB =
3587       EntryBB->splitBasicBlock(EntryBB->getTerminator(), "omp_region.finalize");
3588 
3589   Builder.SetInsertPoint(EntryBB->getTerminator());
3590   emitCommonDirectiveEntry(OMPD, EntryCall, ExitBB, Conditional);
3591 
3592   // generate body
3593   BodyGenCB(/* AllocaIP */ InsertPointTy(),
3594             /* CodeGenIP */ Builder.saveIP());
3595 
3596   // emit exit call and do any needed finalization.
3597   auto FinIP = InsertPointTy(FiniBB, FiniBB->getFirstInsertionPt());
3598   assert(FiniBB->getTerminator()->getNumSuccessors() == 1 &&
3599          FiniBB->getTerminator()->getSuccessor(0) == ExitBB &&
3600          "Unexpected control flow graph state!!");
3601   emitCommonDirectiveExit(OMPD, FinIP, ExitCall, HasFinalize);
3602   assert(FiniBB->getUniquePredecessor()->getUniqueSuccessor() == FiniBB &&
3603          "Unexpected Control Flow State!");
3604   MergeBlockIntoPredecessor(FiniBB);
3605 
3606   // If we are skipping the region of a non conditional, remove the exit
3607   // block, and clear the builder's insertion point.
3608   assert(SplitPos->getParent() == ExitBB &&
3609          "Unexpected Insertion point location!");
3610   auto merged = MergeBlockIntoPredecessor(ExitBB);
3611   BasicBlock *ExitPredBB = SplitPos->getParent();
3612   auto InsertBB = merged ? ExitPredBB : ExitBB;
3613   if (!isa_and_nonnull<BranchInst>(SplitPos))
3614     SplitPos->eraseFromParent();
3615   Builder.SetInsertPoint(InsertBB);
3616 
3617   return Builder.saveIP();
3618 }
3619 
3620 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitCommonDirectiveEntry(
3621     Directive OMPD, Value *EntryCall, BasicBlock *ExitBB, bool Conditional) {
3622   // if nothing to do, Return current insertion point.
3623   if (!Conditional || !EntryCall)
3624     return Builder.saveIP();
3625 
3626   BasicBlock *EntryBB = Builder.GetInsertBlock();
3627   Value *CallBool = Builder.CreateIsNotNull(EntryCall);
3628   auto *ThenBB = BasicBlock::Create(M.getContext(), "omp_region.body");
3629   auto *UI = new UnreachableInst(Builder.getContext(), ThenBB);
3630 
3631   // Emit thenBB and set the Builder's insertion point there for
3632   // body generation next. Place the block after the current block.
3633   Function *CurFn = EntryBB->getParent();
3634   CurFn->insert(std::next(EntryBB->getIterator()), ThenBB);
3635 
3636   // Move Entry branch to end of ThenBB, and replace with conditional
3637   // branch (If-stmt)
3638   Instruction *EntryBBTI = EntryBB->getTerminator();
3639   Builder.CreateCondBr(CallBool, ThenBB, ExitBB);
3640   EntryBBTI->removeFromParent();
3641   Builder.SetInsertPoint(UI);
3642   Builder.Insert(EntryBBTI);
3643   UI->eraseFromParent();
3644   Builder.SetInsertPoint(ThenBB->getTerminator());
3645 
3646   // return an insertion point to ExitBB.
3647   return IRBuilder<>::InsertPoint(ExitBB, ExitBB->getFirstInsertionPt());
3648 }
3649 
3650 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitCommonDirectiveExit(
3651     omp::Directive OMPD, InsertPointTy FinIP, Instruction *ExitCall,
3652     bool HasFinalize) {
3653 
3654   Builder.restoreIP(FinIP);
3655 
3656   // If there is finalization to do, emit it before the exit call
3657   if (HasFinalize) {
3658     assert(!FinalizationStack.empty() &&
3659            "Unexpected finalization stack state!");
3660 
3661     FinalizationInfo Fi = FinalizationStack.pop_back_val();
3662     assert(Fi.DK == OMPD && "Unexpected Directive for Finalization call!");
3663 
3664     Fi.FiniCB(FinIP);
3665 
3666     BasicBlock *FiniBB = FinIP.getBlock();
3667     Instruction *FiniBBTI = FiniBB->getTerminator();
3668 
3669     // set Builder IP for call creation
3670     Builder.SetInsertPoint(FiniBBTI);
3671   }
3672 
3673   if (!ExitCall)
3674     return Builder.saveIP();
3675 
3676   // place the Exitcall as last instruction before Finalization block terminator
3677   ExitCall->removeFromParent();
3678   Builder.Insert(ExitCall);
3679 
3680   return IRBuilder<>::InsertPoint(ExitCall->getParent(),
3681                                   ExitCall->getIterator());
3682 }
3683 
3684 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createCopyinClauseBlocks(
3685     InsertPointTy IP, Value *MasterAddr, Value *PrivateAddr,
3686     llvm::IntegerType *IntPtrTy, bool BranchtoEnd) {
3687   if (!IP.isSet())
3688     return IP;
3689 
3690   IRBuilder<>::InsertPointGuard IPG(Builder);
3691 
3692   // creates the following CFG structure
3693   //	   OMP_Entry : (MasterAddr != PrivateAddr)?
3694   //       F     T
3695   //       |      \
3696   //       |     copin.not.master
3697   //       |      /
3698   //       v     /
3699   //   copyin.not.master.end
3700   //		     |
3701   //         v
3702   //   OMP.Entry.Next
3703 
3704   BasicBlock *OMP_Entry = IP.getBlock();
3705   Function *CurFn = OMP_Entry->getParent();
3706   BasicBlock *CopyBegin =
3707       BasicBlock::Create(M.getContext(), "copyin.not.master", CurFn);
3708   BasicBlock *CopyEnd = nullptr;
3709 
3710   // If entry block is terminated, split to preserve the branch to following
3711   // basic block (i.e. OMP.Entry.Next), otherwise, leave everything as is.
3712   if (isa_and_nonnull<BranchInst>(OMP_Entry->getTerminator())) {
3713     CopyEnd = OMP_Entry->splitBasicBlock(OMP_Entry->getTerminator(),
3714                                          "copyin.not.master.end");
3715     OMP_Entry->getTerminator()->eraseFromParent();
3716   } else {
3717     CopyEnd =
3718         BasicBlock::Create(M.getContext(), "copyin.not.master.end", CurFn);
3719   }
3720 
3721   Builder.SetInsertPoint(OMP_Entry);
3722   Value *MasterPtr = Builder.CreatePtrToInt(MasterAddr, IntPtrTy);
3723   Value *PrivatePtr = Builder.CreatePtrToInt(PrivateAddr, IntPtrTy);
3724   Value *cmp = Builder.CreateICmpNE(MasterPtr, PrivatePtr);
3725   Builder.CreateCondBr(cmp, CopyBegin, CopyEnd);
3726 
3727   Builder.SetInsertPoint(CopyBegin);
3728   if (BranchtoEnd)
3729     Builder.SetInsertPoint(Builder.CreateBr(CopyEnd));
3730 
3731   return Builder.saveIP();
3732 }
3733 
3734 CallInst *OpenMPIRBuilder::createOMPAlloc(const LocationDescription &Loc,
3735                                           Value *Size, Value *Allocator,
3736                                           std::string Name) {
3737   IRBuilder<>::InsertPointGuard IPG(Builder);
3738   Builder.restoreIP(Loc.IP);
3739 
3740   uint32_t SrcLocStrSize;
3741   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3742   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3743   Value *ThreadId = getOrCreateThreadID(Ident);
3744   Value *Args[] = {ThreadId, Size, Allocator};
3745 
3746   Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_alloc);
3747 
3748   return Builder.CreateCall(Fn, Args, Name);
3749 }
3750 
3751 CallInst *OpenMPIRBuilder::createOMPFree(const LocationDescription &Loc,
3752                                          Value *Addr, Value *Allocator,
3753                                          std::string Name) {
3754   IRBuilder<>::InsertPointGuard IPG(Builder);
3755   Builder.restoreIP(Loc.IP);
3756 
3757   uint32_t SrcLocStrSize;
3758   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3759   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3760   Value *ThreadId = getOrCreateThreadID(Ident);
3761   Value *Args[] = {ThreadId, Addr, Allocator};
3762   Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_free);
3763   return Builder.CreateCall(Fn, Args, Name);
3764 }
3765 
3766 CallInst *OpenMPIRBuilder::createOMPInteropInit(
3767     const LocationDescription &Loc, Value *InteropVar,
3768     omp::OMPInteropType InteropType, Value *Device, Value *NumDependences,
3769     Value *DependenceAddress, bool HaveNowaitClause) {
3770   IRBuilder<>::InsertPointGuard IPG(Builder);
3771   Builder.restoreIP(Loc.IP);
3772 
3773   uint32_t SrcLocStrSize;
3774   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3775   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3776   Value *ThreadId = getOrCreateThreadID(Ident);
3777   if (Device == nullptr)
3778     Device = ConstantInt::get(Int32, -1);
3779   Constant *InteropTypeVal = ConstantInt::get(Int32, (int)InteropType);
3780   if (NumDependences == nullptr) {
3781     NumDependences = ConstantInt::get(Int64, 0);
3782     PointerType *PointerTypeVar = Type::getInt8PtrTy(M.getContext());
3783     DependenceAddress = ConstantPointerNull::get(PointerTypeVar);
3784   }
3785   Value *HaveNowaitClauseVal = ConstantInt::get(Int32, HaveNowaitClause);
3786   Value *Args[] = {
3787       Ident,  ThreadId,       InteropVar,        InteropTypeVal,
3788       Device, NumDependences, DependenceAddress, HaveNowaitClauseVal};
3789 
3790   Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___tgt_interop_init);
3791 
3792   return Builder.CreateCall(Fn, Args);
3793 }
3794 
3795 CallInst *OpenMPIRBuilder::createOMPInteropDestroy(
3796     const LocationDescription &Loc, Value *InteropVar, Value *Device,
3797     Value *NumDependences, Value *DependenceAddress, bool HaveNowaitClause) {
3798   IRBuilder<>::InsertPointGuard IPG(Builder);
3799   Builder.restoreIP(Loc.IP);
3800 
3801   uint32_t SrcLocStrSize;
3802   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3803   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3804   Value *ThreadId = getOrCreateThreadID(Ident);
3805   if (Device == nullptr)
3806     Device = ConstantInt::get(Int32, -1);
3807   if (NumDependences == nullptr) {
3808     NumDependences = ConstantInt::get(Int32, 0);
3809     PointerType *PointerTypeVar = Type::getInt8PtrTy(M.getContext());
3810     DependenceAddress = ConstantPointerNull::get(PointerTypeVar);
3811   }
3812   Value *HaveNowaitClauseVal = ConstantInt::get(Int32, HaveNowaitClause);
3813   Value *Args[] = {
3814       Ident,          ThreadId,          InteropVar,         Device,
3815       NumDependences, DependenceAddress, HaveNowaitClauseVal};
3816 
3817   Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___tgt_interop_destroy);
3818 
3819   return Builder.CreateCall(Fn, Args);
3820 }
3821 
3822 CallInst *OpenMPIRBuilder::createOMPInteropUse(const LocationDescription &Loc,
3823                                                Value *InteropVar, Value *Device,
3824                                                Value *NumDependences,
3825                                                Value *DependenceAddress,
3826                                                bool HaveNowaitClause) {
3827   IRBuilder<>::InsertPointGuard IPG(Builder);
3828   Builder.restoreIP(Loc.IP);
3829   uint32_t SrcLocStrSize;
3830   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3831   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3832   Value *ThreadId = getOrCreateThreadID(Ident);
3833   if (Device == nullptr)
3834     Device = ConstantInt::get(Int32, -1);
3835   if (NumDependences == nullptr) {
3836     NumDependences = ConstantInt::get(Int32, 0);
3837     PointerType *PointerTypeVar = Type::getInt8PtrTy(M.getContext());
3838     DependenceAddress = ConstantPointerNull::get(PointerTypeVar);
3839   }
3840   Value *HaveNowaitClauseVal = ConstantInt::get(Int32, HaveNowaitClause);
3841   Value *Args[] = {
3842       Ident,          ThreadId,          InteropVar,         Device,
3843       NumDependences, DependenceAddress, HaveNowaitClauseVal};
3844 
3845   Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___tgt_interop_use);
3846 
3847   return Builder.CreateCall(Fn, Args);
3848 }
3849 
3850 CallInst *OpenMPIRBuilder::createCachedThreadPrivate(
3851     const LocationDescription &Loc, llvm::Value *Pointer,
3852     llvm::ConstantInt *Size, const llvm::Twine &Name) {
3853   IRBuilder<>::InsertPointGuard IPG(Builder);
3854   Builder.restoreIP(Loc.IP);
3855 
3856   uint32_t SrcLocStrSize;
3857   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3858   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3859   Value *ThreadId = getOrCreateThreadID(Ident);
3860   Constant *ThreadPrivateCache =
3861       getOrCreateInternalVariable(Int8PtrPtr, Name.str());
3862   llvm::Value *Args[] = {Ident, ThreadId, Pointer, Size, ThreadPrivateCache};
3863 
3864   Function *Fn =
3865       getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_threadprivate_cached);
3866 
3867   return Builder.CreateCall(Fn, Args);
3868 }
3869 
3870 OpenMPIRBuilder::InsertPointTy
3871 OpenMPIRBuilder::createTargetInit(const LocationDescription &Loc, bool IsSPMD) {
3872   if (!updateToLocation(Loc))
3873     return Loc.IP;
3874 
3875   uint32_t SrcLocStrSize;
3876   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3877   Constant *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3878   ConstantInt *IsSPMDVal = ConstantInt::getSigned(
3879       IntegerType::getInt8Ty(Int8->getContext()),
3880       IsSPMD ? OMP_TGT_EXEC_MODE_SPMD : OMP_TGT_EXEC_MODE_GENERIC);
3881   ConstantInt *UseGenericStateMachine =
3882       ConstantInt::getBool(Int32->getContext(), !IsSPMD);
3883 
3884   Function *Fn = getOrCreateRuntimeFunctionPtr(
3885       omp::RuntimeFunction::OMPRTL___kmpc_target_init);
3886 
3887   CallInst *ThreadKind = Builder.CreateCall(
3888       Fn, {Ident, IsSPMDVal, UseGenericStateMachine});
3889 
3890   Value *ExecUserCode = Builder.CreateICmpEQ(
3891       ThreadKind, ConstantInt::get(ThreadKind->getType(), -1),
3892       "exec_user_code");
3893 
3894   // ThreadKind = __kmpc_target_init(...)
3895   // if (ThreadKind == -1)
3896   //   user_code
3897   // else
3898   //   return;
3899 
3900   auto *UI = Builder.CreateUnreachable();
3901   BasicBlock *CheckBB = UI->getParent();
3902   BasicBlock *UserCodeEntryBB = CheckBB->splitBasicBlock(UI, "user_code.entry");
3903 
3904   BasicBlock *WorkerExitBB = BasicBlock::Create(
3905       CheckBB->getContext(), "worker.exit", CheckBB->getParent());
3906   Builder.SetInsertPoint(WorkerExitBB);
3907   Builder.CreateRetVoid();
3908 
3909   auto *CheckBBTI = CheckBB->getTerminator();
3910   Builder.SetInsertPoint(CheckBBTI);
3911   Builder.CreateCondBr(ExecUserCode, UI->getParent(), WorkerExitBB);
3912 
3913   CheckBBTI->eraseFromParent();
3914   UI->eraseFromParent();
3915 
3916   // Continue in the "user_code" block, see diagram above and in
3917   // openmp/libomptarget/deviceRTLs/common/include/target.h .
3918   return InsertPointTy(UserCodeEntryBB, UserCodeEntryBB->getFirstInsertionPt());
3919 }
3920 
3921 void OpenMPIRBuilder::createTargetDeinit(const LocationDescription &Loc,
3922                                          bool IsSPMD) {
3923   if (!updateToLocation(Loc))
3924     return;
3925 
3926   uint32_t SrcLocStrSize;
3927   Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize);
3928   Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize);
3929   ConstantInt *IsSPMDVal = ConstantInt::getSigned(
3930       IntegerType::getInt8Ty(Int8->getContext()),
3931       IsSPMD ? OMP_TGT_EXEC_MODE_SPMD : OMP_TGT_EXEC_MODE_GENERIC);
3932 
3933   Function *Fn = getOrCreateRuntimeFunctionPtr(
3934       omp::RuntimeFunction::OMPRTL___kmpc_target_deinit);
3935 
3936   Builder.CreateCall(Fn, {Ident, IsSPMDVal});
3937 }
3938 
3939 void OpenMPIRBuilder::setOutlinedTargetRegionFunctionAttributes(
3940     Function *OutlinedFn, int32_t NumTeams, int32_t NumThreads) {
3941   if (Config.isEmbedded()) {
3942     OutlinedFn->setLinkage(GlobalValue::WeakODRLinkage);
3943     // TODO: Determine if DSO local can be set to true.
3944     OutlinedFn->setDSOLocal(false);
3945     OutlinedFn->setVisibility(GlobalValue::ProtectedVisibility);
3946     if (Triple(M.getTargetTriple()).isAMDGCN())
3947       OutlinedFn->setCallingConv(CallingConv::AMDGPU_KERNEL);
3948   }
3949 
3950   if (NumTeams > 0)
3951     OutlinedFn->addFnAttr("omp_target_num_teams", std::to_string(NumTeams));
3952   if (NumThreads > 0)
3953     OutlinedFn->addFnAttr("omp_target_thread_limit",
3954                           std::to_string(NumThreads));
3955 }
3956 
3957 Constant *OpenMPIRBuilder::createOutlinedFunctionID(Function *OutlinedFn,
3958                                                     StringRef EntryFnIDName) {
3959   if (Config.isEmbedded()) {
3960     assert(OutlinedFn && "The outlined function must exist if embedded");
3961     return ConstantExpr::getBitCast(OutlinedFn, Builder.getInt8PtrTy());
3962   }
3963 
3964   return new GlobalVariable(
3965       M, Builder.getInt8Ty(), /*isConstant=*/true, GlobalValue::WeakAnyLinkage,
3966       Constant::getNullValue(Builder.getInt8Ty()), EntryFnIDName);
3967 }
3968 
3969 Constant *OpenMPIRBuilder::createTargetRegionEntryAddr(Function *OutlinedFn,
3970                                                        StringRef EntryFnName) {
3971   if (OutlinedFn)
3972     return OutlinedFn;
3973 
3974   assert(!M.getGlobalVariable(EntryFnName, true) &&
3975          "Named kernel already exists?");
3976   return new GlobalVariable(
3977       M, Builder.getInt8Ty(), /*isConstant=*/true, GlobalValue::InternalLinkage,
3978       Constant::getNullValue(Builder.getInt8Ty()), EntryFnName);
3979 }
3980 
3981 void OpenMPIRBuilder::emitTargetRegionFunction(
3982     OffloadEntriesInfoManager &InfoManager, TargetRegionEntryInfo &EntryInfo,
3983     FunctionGenCallback &GenerateFunctionCallback, int32_t NumTeams,
3984     int32_t NumThreads, bool IsOffloadEntry, Function *&OutlinedFn,
3985     Constant *&OutlinedFnID) {
3986 
3987   SmallString<64> EntryFnName;
3988   InfoManager.getTargetRegionEntryFnName(EntryFnName, EntryInfo);
3989 
3990   OutlinedFn = Config.isEmbedded() || !Config.openMPOffloadMandatory()
3991                    ? GenerateFunctionCallback(EntryFnName)
3992                    : nullptr;
3993 
3994   // If this target outline function is not an offload entry, we don't need to
3995   // register it. This may be in the case of a false if clause, or if there are
3996   // no OpenMP targets.
3997   if (!IsOffloadEntry)
3998     return;
3999 
4000   std::string EntryFnIDName =
4001       Config.isEmbedded()
4002           ? std::string(EntryFnName)
4003           : createPlatformSpecificName({EntryFnName, "region_id"});
4004 
4005   OutlinedFnID = registerTargetRegionFunction(
4006       InfoManager, EntryInfo, OutlinedFn, EntryFnName, EntryFnIDName, NumTeams,
4007       NumThreads);
4008 }
4009 
4010 Constant *OpenMPIRBuilder::registerTargetRegionFunction(
4011     OffloadEntriesInfoManager &InfoManager, TargetRegionEntryInfo &EntryInfo,
4012     Function *OutlinedFn, StringRef EntryFnName, StringRef EntryFnIDName,
4013     int32_t NumTeams, int32_t NumThreads) {
4014   if (OutlinedFn)
4015     setOutlinedTargetRegionFunctionAttributes(OutlinedFn, NumTeams, NumThreads);
4016   auto OutlinedFnID = createOutlinedFunctionID(OutlinedFn, EntryFnIDName);
4017   auto EntryAddr = createTargetRegionEntryAddr(OutlinedFn, EntryFnName);
4018   InfoManager.registerTargetRegionEntryInfo(
4019       EntryInfo, EntryAddr, OutlinedFnID,
4020       OffloadEntriesInfoManager::OMPTargetRegionEntryTargetRegion);
4021   return OutlinedFnID;
4022 }
4023 
4024 std::string OpenMPIRBuilder::getNameWithSeparators(ArrayRef<StringRef> Parts,
4025                                                    StringRef FirstSeparator,
4026                                                    StringRef Separator) {
4027   SmallString<128> Buffer;
4028   llvm::raw_svector_ostream OS(Buffer);
4029   StringRef Sep = FirstSeparator;
4030   for (StringRef Part : Parts) {
4031     OS << Sep << Part;
4032     Sep = Separator;
4033   }
4034   return OS.str().str();
4035 }
4036 
4037 std::string
4038 OpenMPIRBuilder::createPlatformSpecificName(ArrayRef<StringRef> Parts) const {
4039   return OpenMPIRBuilder::getNameWithSeparators(Parts, Config.firstSeparator(),
4040                                                 Config.separator());
4041 }
4042 
4043 GlobalVariable *
4044 OpenMPIRBuilder::getOrCreateInternalVariable(Type *Ty, const StringRef &Name,
4045                                              unsigned AddressSpace) {
4046   auto &Elem = *InternalVars.try_emplace(Name, nullptr).first;
4047   if (Elem.second) {
4048     assert(cast<PointerType>(Elem.second->getType())
4049                ->isOpaqueOrPointeeTypeMatches(Ty) &&
4050            "OMP internal variable has different type than requested");
4051   } else {
4052     // TODO: investigate the appropriate linkage type used for the global
4053     // variable for possibly changing that to internal or private, or maybe
4054     // create different versions of the function for different OMP internal
4055     // variables.
4056     Elem.second = new GlobalVariable(
4057         M, Ty, /*IsConstant=*/false, GlobalValue::CommonLinkage,
4058         Constant::getNullValue(Ty), Elem.first(),
4059         /*InsertBefore=*/nullptr, GlobalValue::NotThreadLocal, AddressSpace);
4060   }
4061 
4062   return cast<GlobalVariable>(&*Elem.second);
4063 }
4064 
4065 Value *OpenMPIRBuilder::getOMPCriticalRegionLock(StringRef CriticalName) {
4066   std::string Prefix = Twine("gomp_critical_user_", CriticalName).str();
4067   std::string Name = getNameWithSeparators({Prefix, "var"}, ".", ".");
4068   return getOrCreateInternalVariable(KmpCriticalNameTy, Name);
4069 }
4070 
4071 GlobalVariable *
4072 OpenMPIRBuilder::createOffloadMaptypes(SmallVectorImpl<uint64_t> &Mappings,
4073                                        std::string VarName) {
4074   llvm::Constant *MaptypesArrayInit =
4075       llvm::ConstantDataArray::get(M.getContext(), Mappings);
4076   auto *MaptypesArrayGlobal = new llvm::GlobalVariable(
4077       M, MaptypesArrayInit->getType(),
4078       /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, MaptypesArrayInit,
4079       VarName);
4080   MaptypesArrayGlobal->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
4081   return MaptypesArrayGlobal;
4082 }
4083 
4084 void OpenMPIRBuilder::createMapperAllocas(const LocationDescription &Loc,
4085                                           InsertPointTy AllocaIP,
4086                                           unsigned NumOperands,
4087                                           struct MapperAllocas &MapperAllocas) {
4088   if (!updateToLocation(Loc))
4089     return;
4090 
4091   auto *ArrI8PtrTy = ArrayType::get(Int8Ptr, NumOperands);
4092   auto *ArrI64Ty = ArrayType::get(Int64, NumOperands);
4093   Builder.restoreIP(AllocaIP);
4094   AllocaInst *ArgsBase = Builder.CreateAlloca(ArrI8PtrTy);
4095   AllocaInst *Args = Builder.CreateAlloca(ArrI8PtrTy);
4096   AllocaInst *ArgSizes = Builder.CreateAlloca(ArrI64Ty);
4097   Builder.restoreIP(Loc.IP);
4098   MapperAllocas.ArgsBase = ArgsBase;
4099   MapperAllocas.Args = Args;
4100   MapperAllocas.ArgSizes = ArgSizes;
4101 }
4102 
4103 void OpenMPIRBuilder::emitMapperCall(const LocationDescription &Loc,
4104                                      Function *MapperFunc, Value *SrcLocInfo,
4105                                      Value *MaptypesArg, Value *MapnamesArg,
4106                                      struct MapperAllocas &MapperAllocas,
4107                                      int64_t DeviceID, unsigned NumOperands) {
4108   if (!updateToLocation(Loc))
4109     return;
4110 
4111   auto *ArrI8PtrTy = ArrayType::get(Int8Ptr, NumOperands);
4112   auto *ArrI64Ty = ArrayType::get(Int64, NumOperands);
4113   Value *ArgsBaseGEP =
4114       Builder.CreateInBoundsGEP(ArrI8PtrTy, MapperAllocas.ArgsBase,
4115                                 {Builder.getInt32(0), Builder.getInt32(0)});
4116   Value *ArgsGEP =
4117       Builder.CreateInBoundsGEP(ArrI8PtrTy, MapperAllocas.Args,
4118                                 {Builder.getInt32(0), Builder.getInt32(0)});
4119   Value *ArgSizesGEP =
4120       Builder.CreateInBoundsGEP(ArrI64Ty, MapperAllocas.ArgSizes,
4121                                 {Builder.getInt32(0), Builder.getInt32(0)});
4122   Value *NullPtr = Constant::getNullValue(Int8Ptr->getPointerTo());
4123   Builder.CreateCall(MapperFunc,
4124                      {SrcLocInfo, Builder.getInt64(DeviceID),
4125                       Builder.getInt32(NumOperands), ArgsBaseGEP, ArgsGEP,
4126                       ArgSizesGEP, MaptypesArg, MapnamesArg, NullPtr});
4127 }
4128 
4129 void OpenMPIRBuilder::emitOffloadingArraysArgument(IRBuilderBase &Builder,
4130                                                    TargetDataRTArgs &RTArgs,
4131                                                    TargetDataInfo &Info,
4132                                                    bool EmitDebug,
4133                                                    bool ForEndCall) {
4134   assert((!ForEndCall || Info.separateBeginEndCalls()) &&
4135          "expected region end call to runtime only when end call is separate");
4136   auto VoidPtrTy = Type::getInt8PtrTy(M.getContext());
4137   auto VoidPtrPtrTy = VoidPtrTy->getPointerTo(0);
4138   auto Int64Ty = Type::getInt64Ty(M.getContext());
4139   auto Int64PtrTy = Type::getInt64PtrTy(M.getContext());
4140 
4141   if (!Info.NumberOfPtrs) {
4142     RTArgs.BasePointersArray = ConstantPointerNull::get(VoidPtrPtrTy);
4143     RTArgs.PointersArray = ConstantPointerNull::get(VoidPtrPtrTy);
4144     RTArgs.SizesArray = ConstantPointerNull::get(Int64PtrTy);
4145     RTArgs.MapTypesArray = ConstantPointerNull::get(Int64PtrTy);
4146     RTArgs.MapNamesArray = ConstantPointerNull::get(VoidPtrPtrTy);
4147     RTArgs.MappersArray = ConstantPointerNull::get(VoidPtrPtrTy);
4148     return;
4149   }
4150 
4151   RTArgs.BasePointersArray = Builder.CreateConstInBoundsGEP2_32(
4152       ArrayType::get(VoidPtrTy, Info.NumberOfPtrs),
4153       Info.RTArgs.BasePointersArray,
4154       /*Idx0=*/0, /*Idx1=*/0);
4155   RTArgs.PointersArray = Builder.CreateConstInBoundsGEP2_32(
4156       ArrayType::get(VoidPtrTy, Info.NumberOfPtrs), Info.RTArgs.PointersArray,
4157       /*Idx0=*/0,
4158       /*Idx1=*/0);
4159   RTArgs.SizesArray = Builder.CreateConstInBoundsGEP2_32(
4160       ArrayType::get(Int64Ty, Info.NumberOfPtrs), Info.RTArgs.SizesArray,
4161       /*Idx0=*/0, /*Idx1=*/0);
4162   RTArgs.MapTypesArray = Builder.CreateConstInBoundsGEP2_32(
4163       ArrayType::get(Int64Ty, Info.NumberOfPtrs),
4164       ForEndCall && Info.RTArgs.MapTypesArrayEnd ? Info.RTArgs.MapTypesArrayEnd
4165                                                  : Info.RTArgs.MapTypesArray,
4166       /*Idx0=*/0,
4167       /*Idx1=*/0);
4168 
4169   // Only emit the mapper information arrays if debug information is
4170   // requested.
4171   if (!EmitDebug)
4172     RTArgs.MapNamesArray = ConstantPointerNull::get(VoidPtrPtrTy);
4173   else
4174     RTArgs.MapNamesArray = Builder.CreateConstInBoundsGEP2_32(
4175         ArrayType::get(VoidPtrTy, Info.NumberOfPtrs), Info.RTArgs.MapNamesArray,
4176         /*Idx0=*/0,
4177         /*Idx1=*/0);
4178   // If there is no user-defined mapper, set the mapper array to nullptr to
4179   // avoid an unnecessary data privatization
4180   if (!Info.HasMapper)
4181     RTArgs.MappersArray = ConstantPointerNull::get(VoidPtrPtrTy);
4182   else
4183     RTArgs.MappersArray =
4184         Builder.CreatePointerCast(Info.RTArgs.MappersArray, VoidPtrPtrTy);
4185 }
4186 
4187 bool OpenMPIRBuilder::checkAndEmitFlushAfterAtomic(
4188     const LocationDescription &Loc, llvm::AtomicOrdering AO, AtomicKind AK) {
4189   assert(!(AO == AtomicOrdering::NotAtomic ||
4190            AO == llvm::AtomicOrdering::Unordered) &&
4191          "Unexpected Atomic Ordering.");
4192 
4193   bool Flush = false;
4194   llvm::AtomicOrdering FlushAO = AtomicOrdering::Monotonic;
4195 
4196   switch (AK) {
4197   case Read:
4198     if (AO == AtomicOrdering::Acquire || AO == AtomicOrdering::AcquireRelease ||
4199         AO == AtomicOrdering::SequentiallyConsistent) {
4200       FlushAO = AtomicOrdering::Acquire;
4201       Flush = true;
4202     }
4203     break;
4204   case Write:
4205   case Compare:
4206   case Update:
4207     if (AO == AtomicOrdering::Release || AO == AtomicOrdering::AcquireRelease ||
4208         AO == AtomicOrdering::SequentiallyConsistent) {
4209       FlushAO = AtomicOrdering::Release;
4210       Flush = true;
4211     }
4212     break;
4213   case Capture:
4214     switch (AO) {
4215     case AtomicOrdering::Acquire:
4216       FlushAO = AtomicOrdering::Acquire;
4217       Flush = true;
4218       break;
4219     case AtomicOrdering::Release:
4220       FlushAO = AtomicOrdering::Release;
4221       Flush = true;
4222       break;
4223     case AtomicOrdering::AcquireRelease:
4224     case AtomicOrdering::SequentiallyConsistent:
4225       FlushAO = AtomicOrdering::AcquireRelease;
4226       Flush = true;
4227       break;
4228     default:
4229       // do nothing - leave silently.
4230       break;
4231     }
4232   }
4233 
4234   if (Flush) {
4235     // Currently Flush RT call still doesn't take memory_ordering, so for when
4236     // that happens, this tries to do the resolution of which atomic ordering
4237     // to use with but issue the flush call
4238     // TODO: pass `FlushAO` after memory ordering support is added
4239     (void)FlushAO;
4240     emitFlush(Loc);
4241   }
4242 
4243   // for AO == AtomicOrdering::Monotonic and  all other case combinations
4244   // do nothing
4245   return Flush;
4246 }
4247 
4248 OpenMPIRBuilder::InsertPointTy
4249 OpenMPIRBuilder::createAtomicRead(const LocationDescription &Loc,
4250                                   AtomicOpValue &X, AtomicOpValue &V,
4251                                   AtomicOrdering AO) {
4252   if (!updateToLocation(Loc))
4253     return Loc.IP;
4254 
4255   Type *XTy = X.Var->getType();
4256   assert(XTy->isPointerTy() && "OMP Atomic expects a pointer to target memory");
4257   Type *XElemTy = X.ElemTy;
4258   assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() ||
4259           XElemTy->isPointerTy()) &&
4260          "OMP atomic read expected a scalar type");
4261 
4262   Value *XRead = nullptr;
4263 
4264   if (XElemTy->isIntegerTy()) {
4265     LoadInst *XLD =
4266         Builder.CreateLoad(XElemTy, X.Var, X.IsVolatile, "omp.atomic.read");
4267     XLD->setAtomic(AO);
4268     XRead = cast<Value>(XLD);
4269   } else {
4270     // We need to bitcast and perform atomic op as integer
4271     unsigned Addrspace = cast<PointerType>(XTy)->getAddressSpace();
4272     IntegerType *IntCastTy =
4273         IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits());
4274     Value *XBCast = Builder.CreateBitCast(
4275         X.Var, IntCastTy->getPointerTo(Addrspace), "atomic.src.int.cast");
4276     LoadInst *XLoad =
4277         Builder.CreateLoad(IntCastTy, XBCast, X.IsVolatile, "omp.atomic.load");
4278     XLoad->setAtomic(AO);
4279     if (XElemTy->isFloatingPointTy()) {
4280       XRead = Builder.CreateBitCast(XLoad, XElemTy, "atomic.flt.cast");
4281     } else {
4282       XRead = Builder.CreateIntToPtr(XLoad, XElemTy, "atomic.ptr.cast");
4283     }
4284   }
4285   checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Read);
4286   Builder.CreateStore(XRead, V.Var, V.IsVolatile);
4287   return Builder.saveIP();
4288 }
4289 
4290 OpenMPIRBuilder::InsertPointTy
4291 OpenMPIRBuilder::createAtomicWrite(const LocationDescription &Loc,
4292                                    AtomicOpValue &X, Value *Expr,
4293                                    AtomicOrdering AO) {
4294   if (!updateToLocation(Loc))
4295     return Loc.IP;
4296 
4297   Type *XTy = X.Var->getType();
4298   assert(XTy->isPointerTy() && "OMP Atomic expects a pointer to target memory");
4299   Type *XElemTy = X.ElemTy;
4300   assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() ||
4301           XElemTy->isPointerTy()) &&
4302          "OMP atomic write expected a scalar type");
4303 
4304   if (XElemTy->isIntegerTy()) {
4305     StoreInst *XSt = Builder.CreateStore(Expr, X.Var, X.IsVolatile);
4306     XSt->setAtomic(AO);
4307   } else {
4308     // We need to bitcast and perform atomic op as integers
4309     unsigned Addrspace = cast<PointerType>(XTy)->getAddressSpace();
4310     IntegerType *IntCastTy =
4311         IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits());
4312     Value *XBCast = Builder.CreateBitCast(
4313         X.Var, IntCastTy->getPointerTo(Addrspace), "atomic.dst.int.cast");
4314     Value *ExprCast =
4315         Builder.CreateBitCast(Expr, IntCastTy, "atomic.src.int.cast");
4316     StoreInst *XSt = Builder.CreateStore(ExprCast, XBCast, X.IsVolatile);
4317     XSt->setAtomic(AO);
4318   }
4319 
4320   checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Write);
4321   return Builder.saveIP();
4322 }
4323 
4324 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicUpdate(
4325     const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X,
4326     Value *Expr, AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp,
4327     AtomicUpdateCallbackTy &UpdateOp, bool IsXBinopExpr) {
4328   assert(!isConflictIP(Loc.IP, AllocaIP) && "IPs must not be ambiguous");
4329   if (!updateToLocation(Loc))
4330     return Loc.IP;
4331 
4332   LLVM_DEBUG({
4333     Type *XTy = X.Var->getType();
4334     assert(XTy->isPointerTy() &&
4335            "OMP Atomic expects a pointer to target memory");
4336     Type *XElemTy = X.ElemTy;
4337     assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() ||
4338             XElemTy->isPointerTy()) &&
4339            "OMP atomic update expected a scalar type");
4340     assert((RMWOp != AtomicRMWInst::Max) && (RMWOp != AtomicRMWInst::Min) &&
4341            (RMWOp != AtomicRMWInst::UMax) && (RMWOp != AtomicRMWInst::UMin) &&
4342            "OpenMP atomic does not support LT or GT operations");
4343   });
4344 
4345   emitAtomicUpdate(AllocaIP, X.Var, X.ElemTy, Expr, AO, RMWOp, UpdateOp,
4346                    X.IsVolatile, IsXBinopExpr);
4347   checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Update);
4348   return Builder.saveIP();
4349 }
4350 
4351 // FIXME: Duplicating AtomicExpand
4352 Value *OpenMPIRBuilder::emitRMWOpAsInstruction(Value *Src1, Value *Src2,
4353                                                AtomicRMWInst::BinOp RMWOp) {
4354   switch (RMWOp) {
4355   case AtomicRMWInst::Add:
4356     return Builder.CreateAdd(Src1, Src2);
4357   case AtomicRMWInst::Sub:
4358     return Builder.CreateSub(Src1, Src2);
4359   case AtomicRMWInst::And:
4360     return Builder.CreateAnd(Src1, Src2);
4361   case AtomicRMWInst::Nand:
4362     return Builder.CreateNeg(Builder.CreateAnd(Src1, Src2));
4363   case AtomicRMWInst::Or:
4364     return Builder.CreateOr(Src1, Src2);
4365   case AtomicRMWInst::Xor:
4366     return Builder.CreateXor(Src1, Src2);
4367   case AtomicRMWInst::Xchg:
4368   case AtomicRMWInst::FAdd:
4369   case AtomicRMWInst::FSub:
4370   case AtomicRMWInst::BAD_BINOP:
4371   case AtomicRMWInst::Max:
4372   case AtomicRMWInst::Min:
4373   case AtomicRMWInst::UMax:
4374   case AtomicRMWInst::UMin:
4375   case AtomicRMWInst::FMax:
4376   case AtomicRMWInst::FMin:
4377   case AtomicRMWInst::UIncWrap:
4378   case AtomicRMWInst::UDecWrap:
4379     llvm_unreachable("Unsupported atomic update operation");
4380   }
4381   llvm_unreachable("Unsupported atomic update operation");
4382 }
4383 
4384 std::pair<Value *, Value *> OpenMPIRBuilder::emitAtomicUpdate(
4385     InsertPointTy AllocaIP, Value *X, Type *XElemTy, Value *Expr,
4386     AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp,
4387     AtomicUpdateCallbackTy &UpdateOp, bool VolatileX, bool IsXBinopExpr) {
4388   // TODO: handle the case where XElemTy is not byte-sized or not a power of 2
4389   // or a complex datatype.
4390   bool emitRMWOp = false;
4391   switch (RMWOp) {
4392   case AtomicRMWInst::Add:
4393   case AtomicRMWInst::And:
4394   case AtomicRMWInst::Nand:
4395   case AtomicRMWInst::Or:
4396   case AtomicRMWInst::Xor:
4397   case AtomicRMWInst::Xchg:
4398     emitRMWOp = XElemTy;
4399     break;
4400   case AtomicRMWInst::Sub:
4401     emitRMWOp = (IsXBinopExpr && XElemTy);
4402     break;
4403   default:
4404     emitRMWOp = false;
4405   }
4406   emitRMWOp &= XElemTy->isIntegerTy();
4407 
4408   std::pair<Value *, Value *> Res;
4409   if (emitRMWOp) {
4410     Res.first = Builder.CreateAtomicRMW(RMWOp, X, Expr, llvm::MaybeAlign(), AO);
4411     // not needed except in case of postfix captures. Generate anyway for
4412     // consistency with the else part. Will be removed with any DCE pass.
4413     // AtomicRMWInst::Xchg does not have a coressponding instruction.
4414     if (RMWOp == AtomicRMWInst::Xchg)
4415       Res.second = Res.first;
4416     else
4417       Res.second = emitRMWOpAsInstruction(Res.first, Expr, RMWOp);
4418   } else {
4419     unsigned Addrspace = cast<PointerType>(X->getType())->getAddressSpace();
4420     IntegerType *IntCastTy =
4421         IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits());
4422     Value *XBCast =
4423         Builder.CreateBitCast(X, IntCastTy->getPointerTo(Addrspace));
4424     LoadInst *OldVal =
4425         Builder.CreateLoad(IntCastTy, XBCast, X->getName() + ".atomic.load");
4426     OldVal->setAtomic(AO);
4427     // CurBB
4428     // |     /---\
4429 		// ContBB    |
4430     // |     \---/
4431     // ExitBB
4432     BasicBlock *CurBB = Builder.GetInsertBlock();
4433     Instruction *CurBBTI = CurBB->getTerminator();
4434     CurBBTI = CurBBTI ? CurBBTI : Builder.CreateUnreachable();
4435     BasicBlock *ExitBB =
4436         CurBB->splitBasicBlock(CurBBTI, X->getName() + ".atomic.exit");
4437     BasicBlock *ContBB = CurBB->splitBasicBlock(CurBB->getTerminator(),
4438                                                 X->getName() + ".atomic.cont");
4439     ContBB->getTerminator()->eraseFromParent();
4440     Builder.restoreIP(AllocaIP);
4441     AllocaInst *NewAtomicAddr = Builder.CreateAlloca(XElemTy);
4442     NewAtomicAddr->setName(X->getName() + "x.new.val");
4443     Builder.SetInsertPoint(ContBB);
4444     llvm::PHINode *PHI = Builder.CreatePHI(OldVal->getType(), 2);
4445     PHI->addIncoming(OldVal, CurBB);
4446     IntegerType *NewAtomicCastTy =
4447         IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits());
4448     bool IsIntTy = XElemTy->isIntegerTy();
4449     Value *NewAtomicIntAddr =
4450         (IsIntTy)
4451             ? NewAtomicAddr
4452             : Builder.CreateBitCast(NewAtomicAddr,
4453                                     NewAtomicCastTy->getPointerTo(Addrspace));
4454     Value *OldExprVal = PHI;
4455     if (!IsIntTy) {
4456       if (XElemTy->isFloatingPointTy()) {
4457         OldExprVal = Builder.CreateBitCast(PHI, XElemTy,
4458                                            X->getName() + ".atomic.fltCast");
4459       } else {
4460         OldExprVal = Builder.CreateIntToPtr(PHI, XElemTy,
4461                                             X->getName() + ".atomic.ptrCast");
4462       }
4463     }
4464 
4465     Value *Upd = UpdateOp(OldExprVal, Builder);
4466     Builder.CreateStore(Upd, NewAtomicAddr);
4467     LoadInst *DesiredVal = Builder.CreateLoad(IntCastTy, NewAtomicIntAddr);
4468     Value *XAddr =
4469         (IsIntTy)
4470             ? X
4471             : Builder.CreateBitCast(X, IntCastTy->getPointerTo(Addrspace));
4472     AtomicOrdering Failure =
4473         llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
4474     AtomicCmpXchgInst *Result = Builder.CreateAtomicCmpXchg(
4475         XAddr, PHI, DesiredVal, llvm::MaybeAlign(), AO, Failure);
4476     Result->setVolatile(VolatileX);
4477     Value *PreviousVal = Builder.CreateExtractValue(Result, /*Idxs=*/0);
4478     Value *SuccessFailureVal = Builder.CreateExtractValue(Result, /*Idxs=*/1);
4479     PHI->addIncoming(PreviousVal, Builder.GetInsertBlock());
4480     Builder.CreateCondBr(SuccessFailureVal, ExitBB, ContBB);
4481 
4482     Res.first = OldExprVal;
4483     Res.second = Upd;
4484 
4485     // set Insertion point in exit block
4486     if (UnreachableInst *ExitTI =
4487             dyn_cast<UnreachableInst>(ExitBB->getTerminator())) {
4488       CurBBTI->eraseFromParent();
4489       Builder.SetInsertPoint(ExitBB);
4490     } else {
4491       Builder.SetInsertPoint(ExitTI);
4492     }
4493   }
4494 
4495   return Res;
4496 }
4497 
4498 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicCapture(
4499     const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X,
4500     AtomicOpValue &V, Value *Expr, AtomicOrdering AO,
4501     AtomicRMWInst::BinOp RMWOp, AtomicUpdateCallbackTy &UpdateOp,
4502     bool UpdateExpr, bool IsPostfixUpdate, bool IsXBinopExpr) {
4503   if (!updateToLocation(Loc))
4504     return Loc.IP;
4505 
4506   LLVM_DEBUG({
4507     Type *XTy = X.Var->getType();
4508     assert(XTy->isPointerTy() &&
4509            "OMP Atomic expects a pointer to target memory");
4510     Type *XElemTy = X.ElemTy;
4511     assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() ||
4512             XElemTy->isPointerTy()) &&
4513            "OMP atomic capture expected a scalar type");
4514     assert((RMWOp != AtomicRMWInst::Max) && (RMWOp != AtomicRMWInst::Min) &&
4515            "OpenMP atomic does not support LT or GT operations");
4516   });
4517 
4518   // If UpdateExpr is 'x' updated with some `expr` not based on 'x',
4519   // 'x' is simply atomically rewritten with 'expr'.
4520   AtomicRMWInst::BinOp AtomicOp = (UpdateExpr ? RMWOp : AtomicRMWInst::Xchg);
4521   std::pair<Value *, Value *> Result =
4522       emitAtomicUpdate(AllocaIP, X.Var, X.ElemTy, Expr, AO, AtomicOp, UpdateOp,
4523                        X.IsVolatile, IsXBinopExpr);
4524 
4525   Value *CapturedVal = (IsPostfixUpdate ? Result.first : Result.second);
4526   Builder.CreateStore(CapturedVal, V.Var, V.IsVolatile);
4527 
4528   checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Capture);
4529   return Builder.saveIP();
4530 }
4531 
4532 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicCompare(
4533     const LocationDescription &Loc, AtomicOpValue &X, AtomicOpValue &V,
4534     AtomicOpValue &R, Value *E, Value *D, AtomicOrdering AO,
4535     omp::OMPAtomicCompareOp Op, bool IsXBinopExpr, bool IsPostfixUpdate,
4536     bool IsFailOnly) {
4537 
4538   if (!updateToLocation(Loc))
4539     return Loc.IP;
4540 
4541   assert(X.Var->getType()->isPointerTy() &&
4542          "OMP atomic expects a pointer to target memory");
4543   // compare capture
4544   if (V.Var) {
4545     assert(V.Var->getType()->isPointerTy() && "v.var must be of pointer type");
4546     assert(V.ElemTy == X.ElemTy && "x and v must be of same type");
4547   }
4548 
4549   bool IsInteger = E->getType()->isIntegerTy();
4550 
4551   if (Op == OMPAtomicCompareOp::EQ) {
4552     AtomicOrdering Failure = AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
4553     AtomicCmpXchgInst *Result = nullptr;
4554     if (!IsInteger) {
4555       unsigned Addrspace =
4556           cast<PointerType>(X.Var->getType())->getAddressSpace();
4557       IntegerType *IntCastTy =
4558           IntegerType::get(M.getContext(), X.ElemTy->getScalarSizeInBits());
4559       Value *XBCast =
4560           Builder.CreateBitCast(X.Var, IntCastTy->getPointerTo(Addrspace));
4561       Value *EBCast = Builder.CreateBitCast(E, IntCastTy);
4562       Value *DBCast = Builder.CreateBitCast(D, IntCastTy);
4563       Result = Builder.CreateAtomicCmpXchg(XBCast, EBCast, DBCast, MaybeAlign(),
4564                                            AO, Failure);
4565     } else {
4566       Result =
4567           Builder.CreateAtomicCmpXchg(X.Var, E, D, MaybeAlign(), AO, Failure);
4568     }
4569 
4570     if (V.Var) {
4571       Value *OldValue = Builder.CreateExtractValue(Result, /*Idxs=*/0);
4572       if (!IsInteger)
4573         OldValue = Builder.CreateBitCast(OldValue, X.ElemTy);
4574       assert(OldValue->getType() == V.ElemTy &&
4575              "OldValue and V must be of same type");
4576       if (IsPostfixUpdate) {
4577         Builder.CreateStore(OldValue, V.Var, V.IsVolatile);
4578       } else {
4579         Value *SuccessOrFail = Builder.CreateExtractValue(Result, /*Idxs=*/1);
4580         if (IsFailOnly) {
4581           // CurBB----
4582           //   |     |
4583           //   v     |
4584           // ContBB  |
4585           //   |     |
4586           //   v     |
4587           // ExitBB <-
4588           //
4589           // where ContBB only contains the store of old value to 'v'.
4590           BasicBlock *CurBB = Builder.GetInsertBlock();
4591           Instruction *CurBBTI = CurBB->getTerminator();
4592           CurBBTI = CurBBTI ? CurBBTI : Builder.CreateUnreachable();
4593           BasicBlock *ExitBB = CurBB->splitBasicBlock(
4594               CurBBTI, X.Var->getName() + ".atomic.exit");
4595           BasicBlock *ContBB = CurBB->splitBasicBlock(
4596               CurBB->getTerminator(), X.Var->getName() + ".atomic.cont");
4597           ContBB->getTerminator()->eraseFromParent();
4598           CurBB->getTerminator()->eraseFromParent();
4599 
4600           Builder.CreateCondBr(SuccessOrFail, ExitBB, ContBB);
4601 
4602           Builder.SetInsertPoint(ContBB);
4603           Builder.CreateStore(OldValue, V.Var);
4604           Builder.CreateBr(ExitBB);
4605 
4606           if (UnreachableInst *ExitTI =
4607                   dyn_cast<UnreachableInst>(ExitBB->getTerminator())) {
4608             CurBBTI->eraseFromParent();
4609             Builder.SetInsertPoint(ExitBB);
4610           } else {
4611             Builder.SetInsertPoint(ExitTI);
4612           }
4613         } else {
4614           Value *CapturedValue =
4615               Builder.CreateSelect(SuccessOrFail, E, OldValue);
4616           Builder.CreateStore(CapturedValue, V.Var, V.IsVolatile);
4617         }
4618       }
4619     }
4620     // The comparison result has to be stored.
4621     if (R.Var) {
4622       assert(R.Var->getType()->isPointerTy() &&
4623              "r.var must be of pointer type");
4624       assert(R.ElemTy->isIntegerTy() && "r must be of integral type");
4625 
4626       Value *SuccessFailureVal = Builder.CreateExtractValue(Result, /*Idxs=*/1);
4627       Value *ResultCast = R.IsSigned
4628                               ? Builder.CreateSExt(SuccessFailureVal, R.ElemTy)
4629                               : Builder.CreateZExt(SuccessFailureVal, R.ElemTy);
4630       Builder.CreateStore(ResultCast, R.Var, R.IsVolatile);
4631     }
4632   } else {
4633     assert((Op == OMPAtomicCompareOp::MAX || Op == OMPAtomicCompareOp::MIN) &&
4634            "Op should be either max or min at this point");
4635     assert(!IsFailOnly && "IsFailOnly is only valid when the comparison is ==");
4636 
4637     // Reverse the ordop as the OpenMP forms are different from LLVM forms.
4638     // Let's take max as example.
4639     // OpenMP form:
4640     // x = x > expr ? expr : x;
4641     // LLVM form:
4642     // *ptr = *ptr > val ? *ptr : val;
4643     // We need to transform to LLVM form.
4644     // x = x <= expr ? x : expr;
4645     AtomicRMWInst::BinOp NewOp;
4646     if (IsXBinopExpr) {
4647       if (IsInteger) {
4648         if (X.IsSigned)
4649           NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::Min
4650                                                 : AtomicRMWInst::Max;
4651         else
4652           NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::UMin
4653                                                 : AtomicRMWInst::UMax;
4654       } else {
4655         NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::FMin
4656                                               : AtomicRMWInst::FMax;
4657       }
4658     } else {
4659       if (IsInteger) {
4660         if (X.IsSigned)
4661           NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::Max
4662                                                 : AtomicRMWInst::Min;
4663         else
4664           NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::UMax
4665                                                 : AtomicRMWInst::UMin;
4666       } else {
4667         NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::FMax
4668                                               : AtomicRMWInst::FMin;
4669       }
4670     }
4671 
4672     AtomicRMWInst *OldValue =
4673         Builder.CreateAtomicRMW(NewOp, X.Var, E, MaybeAlign(), AO);
4674     if (V.Var) {
4675       Value *CapturedValue = nullptr;
4676       if (IsPostfixUpdate) {
4677         CapturedValue = OldValue;
4678       } else {
4679         CmpInst::Predicate Pred;
4680         switch (NewOp) {
4681         case AtomicRMWInst::Max:
4682           Pred = CmpInst::ICMP_SGT;
4683           break;
4684         case AtomicRMWInst::UMax:
4685           Pred = CmpInst::ICMP_UGT;
4686           break;
4687         case AtomicRMWInst::FMax:
4688           Pred = CmpInst::FCMP_OGT;
4689           break;
4690         case AtomicRMWInst::Min:
4691           Pred = CmpInst::ICMP_SLT;
4692           break;
4693         case AtomicRMWInst::UMin:
4694           Pred = CmpInst::ICMP_ULT;
4695           break;
4696         case AtomicRMWInst::FMin:
4697           Pred = CmpInst::FCMP_OLT;
4698           break;
4699         default:
4700           llvm_unreachable("unexpected comparison op");
4701         }
4702         Value *NonAtomicCmp = Builder.CreateCmp(Pred, OldValue, E);
4703         CapturedValue = Builder.CreateSelect(NonAtomicCmp, E, OldValue);
4704       }
4705       Builder.CreateStore(CapturedValue, V.Var, V.IsVolatile);
4706     }
4707   }
4708 
4709   checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Compare);
4710 
4711   return Builder.saveIP();
4712 }
4713 
4714 GlobalVariable *
4715 OpenMPIRBuilder::createOffloadMapnames(SmallVectorImpl<llvm::Constant *> &Names,
4716                                        std::string VarName) {
4717   llvm::Constant *MapNamesArrayInit = llvm::ConstantArray::get(
4718       llvm::ArrayType::get(
4719           llvm::Type::getInt8Ty(M.getContext())->getPointerTo(), Names.size()),
4720       Names);
4721   auto *MapNamesArrayGlobal = new llvm::GlobalVariable(
4722       M, MapNamesArrayInit->getType(),
4723       /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, MapNamesArrayInit,
4724       VarName);
4725   return MapNamesArrayGlobal;
4726 }
4727 
4728 // Create all simple and struct types exposed by the runtime and remember
4729 // the llvm::PointerTypes of them for easy access later.
4730 void OpenMPIRBuilder::initializeTypes(Module &M) {
4731   LLVMContext &Ctx = M.getContext();
4732   StructType *T;
4733 #define OMP_TYPE(VarName, InitValue) VarName = InitValue;
4734 #define OMP_ARRAY_TYPE(VarName, ElemTy, ArraySize)                             \
4735   VarName##Ty = ArrayType::get(ElemTy, ArraySize);                             \
4736   VarName##PtrTy = PointerType::getUnqual(VarName##Ty);
4737 #define OMP_FUNCTION_TYPE(VarName, IsVarArg, ReturnType, ...)                  \
4738   VarName = FunctionType::get(ReturnType, {__VA_ARGS__}, IsVarArg);            \
4739   VarName##Ptr = PointerType::getUnqual(VarName);
4740 #define OMP_STRUCT_TYPE(VarName, StructName, Packed, ...)                      \
4741   T = StructType::getTypeByName(Ctx, StructName);                              \
4742   if (!T)                                                                      \
4743     T = StructType::create(Ctx, {__VA_ARGS__}, StructName, Packed);            \
4744   VarName = T;                                                                 \
4745   VarName##Ptr = PointerType::getUnqual(T);
4746 #include "llvm/Frontend/OpenMP/OMPKinds.def"
4747 }
4748 
4749 void OpenMPIRBuilder::OutlineInfo::collectBlocks(
4750     SmallPtrSetImpl<BasicBlock *> &BlockSet,
4751     SmallVectorImpl<BasicBlock *> &BlockVector) {
4752   SmallVector<BasicBlock *, 32> Worklist;
4753   BlockSet.insert(EntryBB);
4754   BlockSet.insert(ExitBB);
4755 
4756   Worklist.push_back(EntryBB);
4757   while (!Worklist.empty()) {
4758     BasicBlock *BB = Worklist.pop_back_val();
4759     BlockVector.push_back(BB);
4760     for (BasicBlock *SuccBB : successors(BB))
4761       if (BlockSet.insert(SuccBB).second)
4762         Worklist.push_back(SuccBB);
4763   }
4764 }
4765 
4766 void OpenMPIRBuilder::createOffloadEntry(Constant *ID, Constant *Addr,
4767                                          uint64_t Size, int32_t Flags,
4768                                          GlobalValue::LinkageTypes) {
4769   if (!Config.isTargetCodegen()) {
4770     emitOffloadingEntry(ID, Addr->getName(), Size, Flags);
4771     return;
4772   }
4773   // TODO: Add support for global variables on the device after declare target
4774   // support.
4775   Function *Fn = dyn_cast<Function>(Addr);
4776   if (!Fn)
4777     return;
4778 
4779   Module &M = *(Fn->getParent());
4780   LLVMContext &Ctx = M.getContext();
4781 
4782   // Get "nvvm.annotations" metadata node.
4783   NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations");
4784 
4785   Metadata *MDVals[] = {
4786       ConstantAsMetadata::get(Fn), MDString::get(Ctx, "kernel"),
4787       ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(Ctx), 1))};
4788   // Append metadata to nvvm.annotations.
4789   MD->addOperand(MDNode::get(Ctx, MDVals));
4790 
4791   // Add a function attribute for the kernel.
4792   Fn->addFnAttr(Attribute::get(Ctx, "kernel"));
4793 }
4794 
4795 // We only generate metadata for function that contain target regions.
4796 void OpenMPIRBuilder::createOffloadEntriesAndInfoMetadata(
4797     OffloadEntriesInfoManager &OffloadEntriesInfoManager,
4798     EmitMetadataErrorReportFunctionTy &ErrorFn) {
4799 
4800   // If there are no entries, we don't need to do anything.
4801   if (OffloadEntriesInfoManager.empty())
4802     return;
4803 
4804   LLVMContext &C = M.getContext();
4805   SmallVector<std::pair<const OffloadEntriesInfoManager::OffloadEntryInfo *,
4806                         TargetRegionEntryInfo>,
4807               16>
4808       OrderedEntries(OffloadEntriesInfoManager.size());
4809 
4810   // Auxiliary methods to create metadata values and strings.
4811   auto &&GetMDInt = [this](unsigned V) {
4812     return ConstantAsMetadata::get(ConstantInt::get(Builder.getInt32Ty(), V));
4813   };
4814 
4815   auto &&GetMDString = [&C](StringRef V) { return MDString::get(C, V); };
4816 
4817   // Create the offloading info metadata node.
4818   NamedMDNode *MD = M.getOrInsertNamedMetadata("omp_offload.info");
4819   auto &&TargetRegionMetadataEmitter =
4820       [&C, MD, &OrderedEntries, &GetMDInt, &GetMDString](
4821           const TargetRegionEntryInfo &EntryInfo,
4822           const OffloadEntriesInfoManager::OffloadEntryInfoTargetRegion &E) {
4823         // Generate metadata for target regions. Each entry of this metadata
4824         // contains:
4825         // - Entry 0 -> Kind of this type of metadata (0).
4826         // - Entry 1 -> Device ID of the file where the entry was identified.
4827         // - Entry 2 -> File ID of the file where the entry was identified.
4828         // - Entry 3 -> Mangled name of the function where the entry was
4829         // identified.
4830         // - Entry 4 -> Line in the file where the entry was identified.
4831         // - Entry 5 -> Count of regions at this DeviceID/FilesID/Line.
4832         // - Entry 6 -> Order the entry was created.
4833         // The first element of the metadata node is the kind.
4834         Metadata *Ops[] = {
4835             GetMDInt(E.getKind()),      GetMDInt(EntryInfo.DeviceID),
4836             GetMDInt(EntryInfo.FileID), GetMDString(EntryInfo.ParentName),
4837             GetMDInt(EntryInfo.Line),   GetMDInt(EntryInfo.Count),
4838             GetMDInt(E.getOrder())};
4839 
4840         // Save this entry in the right position of the ordered entries array.
4841         OrderedEntries[E.getOrder()] = std::make_pair(&E, EntryInfo);
4842 
4843         // Add metadata to the named metadata node.
4844         MD->addOperand(MDNode::get(C, Ops));
4845       };
4846 
4847   OffloadEntriesInfoManager.actOnTargetRegionEntriesInfo(
4848       TargetRegionMetadataEmitter);
4849 
4850   // Create function that emits metadata for each device global variable entry;
4851   auto &&DeviceGlobalVarMetadataEmitter =
4852       [&C, &OrderedEntries, &GetMDInt, &GetMDString, MD](
4853           StringRef MangledName,
4854           const OffloadEntriesInfoManager::OffloadEntryInfoDeviceGlobalVar &E) {
4855         // Generate metadata for global variables. Each entry of this metadata
4856         // contains:
4857         // - Entry 0 -> Kind of this type of metadata (1).
4858         // - Entry 1 -> Mangled name of the variable.
4859         // - Entry 2 -> Declare target kind.
4860         // - Entry 3 -> Order the entry was created.
4861         // The first element of the metadata node is the kind.
4862         Metadata *Ops[] = {GetMDInt(E.getKind()), GetMDString(MangledName),
4863                            GetMDInt(E.getFlags()), GetMDInt(E.getOrder())};
4864 
4865         // Save this entry in the right position of the ordered entries array.
4866         TargetRegionEntryInfo varInfo(MangledName, 0, 0, 0);
4867         OrderedEntries[E.getOrder()] = std::make_pair(&E, varInfo);
4868 
4869         // Add metadata to the named metadata node.
4870         MD->addOperand(MDNode::get(C, Ops));
4871       };
4872 
4873   OffloadEntriesInfoManager.actOnDeviceGlobalVarEntriesInfo(
4874       DeviceGlobalVarMetadataEmitter);
4875 
4876   for (const auto &E : OrderedEntries) {
4877     assert(E.first && "All ordered entries must exist!");
4878     if (const auto *CE =
4879             dyn_cast<OffloadEntriesInfoManager::OffloadEntryInfoTargetRegion>(
4880                 E.first)) {
4881       if (!CE->getID() || !CE->getAddress()) {
4882         // Do not blame the entry if the parent funtion is not emitted.
4883         TargetRegionEntryInfo EntryInfo = E.second;
4884         StringRef FnName = EntryInfo.ParentName;
4885         if (!M.getNamedValue(FnName))
4886           continue;
4887         ErrorFn(EMIT_MD_TARGET_REGION_ERROR, EntryInfo);
4888         continue;
4889       }
4890       createOffloadEntry(CE->getID(), CE->getAddress(),
4891                          /*Size=*/0, CE->getFlags(),
4892                          GlobalValue::WeakAnyLinkage);
4893     } else if (const auto *CE = dyn_cast<
4894                    OffloadEntriesInfoManager::OffloadEntryInfoDeviceGlobalVar>(
4895                    E.first)) {
4896       OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind Flags =
4897           static_cast<OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind>(
4898               CE->getFlags());
4899       switch (Flags) {
4900       case OffloadEntriesInfoManager::OMPTargetGlobalVarEntryTo: {
4901         if (Config.isEmbedded() && Config.hasRequiresUnifiedSharedMemory())
4902           continue;
4903         if (!CE->getAddress()) {
4904           ErrorFn(EMIT_MD_DECLARE_TARGET_ERROR, E.second);
4905           continue;
4906         }
4907         // The vaiable has no definition - no need to add the entry.
4908         if (CE->getVarSize() == 0)
4909           continue;
4910         break;
4911       }
4912       case OffloadEntriesInfoManager::OMPTargetGlobalVarEntryLink:
4913         assert(((Config.isEmbedded() && !CE->getAddress()) ||
4914                 (!Config.isEmbedded() && CE->getAddress())) &&
4915                "Declaret target link address is set.");
4916         if (Config.isEmbedded())
4917           continue;
4918         if (!CE->getAddress()) {
4919           ErrorFn(EMIT_MD_GLOBAL_VAR_LINK_ERROR, TargetRegionEntryInfo());
4920           continue;
4921         }
4922         break;
4923       }
4924 
4925       // Hidden or internal symbols on the device are not externally visible.
4926       // We should not attempt to register them by creating an offloading
4927       // entry.
4928       if (auto *GV = dyn_cast<GlobalValue>(CE->getAddress()))
4929         if (GV->hasLocalLinkage() || GV->hasHiddenVisibility())
4930           continue;
4931 
4932       createOffloadEntry(CE->getAddress(), CE->getAddress(), CE->getVarSize(),
4933                          Flags, CE->getLinkage());
4934 
4935     } else {
4936       llvm_unreachable("Unsupported entry kind.");
4937     }
4938   }
4939 }
4940 
4941 void TargetRegionEntryInfo::getTargetRegionEntryFnName(
4942     SmallVectorImpl<char> &Name, StringRef ParentName, unsigned DeviceID,
4943     unsigned FileID, unsigned Line, unsigned Count) {
4944   raw_svector_ostream OS(Name);
4945   OS << "__omp_offloading" << llvm::format("_%x", DeviceID)
4946      << llvm::format("_%x_", FileID) << ParentName << "_l" << Line;
4947   if (Count)
4948     OS << "_" << Count;
4949 }
4950 
4951 void OffloadEntriesInfoManager::getTargetRegionEntryFnName(
4952     SmallVectorImpl<char> &Name, const TargetRegionEntryInfo &EntryInfo) {
4953   unsigned NewCount = getTargetRegionEntryInfoCount(EntryInfo);
4954   TargetRegionEntryInfo::getTargetRegionEntryFnName(
4955       Name, EntryInfo.ParentName, EntryInfo.DeviceID, EntryInfo.FileID,
4956       EntryInfo.Line, NewCount);
4957 }
4958 
4959 /// Loads all the offload entries information from the host IR
4960 /// metadata.
4961 void OpenMPIRBuilder::loadOffloadInfoMetadata(
4962     Module &M, OffloadEntriesInfoManager &OffloadEntriesInfoManager) {
4963   // If we are in target mode, load the metadata from the host IR. This code has
4964   // to match the metadata creation in createOffloadEntriesAndInfoMetadata().
4965 
4966   NamedMDNode *MD = M.getNamedMetadata(ompOffloadInfoName);
4967   if (!MD)
4968     return;
4969 
4970   for (MDNode *MN : MD->operands()) {
4971     auto &&GetMDInt = [MN](unsigned Idx) {
4972       auto *V = cast<ConstantAsMetadata>(MN->getOperand(Idx));
4973       return cast<ConstantInt>(V->getValue())->getZExtValue();
4974     };
4975 
4976     auto &&GetMDString = [MN](unsigned Idx) {
4977       auto *V = cast<MDString>(MN->getOperand(Idx));
4978       return V->getString();
4979     };
4980 
4981     switch (GetMDInt(0)) {
4982     default:
4983       llvm_unreachable("Unexpected metadata!");
4984       break;
4985     case OffloadEntriesInfoManager::OffloadEntryInfo::
4986         OffloadingEntryInfoTargetRegion: {
4987       TargetRegionEntryInfo EntryInfo(/*ParentName=*/GetMDString(3),
4988                                       /*DeviceID=*/GetMDInt(1),
4989                                       /*FileID=*/GetMDInt(2),
4990                                       /*Line=*/GetMDInt(4),
4991                                       /*Count=*/GetMDInt(5));
4992       OffloadEntriesInfoManager.initializeTargetRegionEntryInfo(
4993           EntryInfo, /*Order=*/GetMDInt(6));
4994       break;
4995     }
4996     case OffloadEntriesInfoManager::OffloadEntryInfo::
4997         OffloadingEntryInfoDeviceGlobalVar:
4998       OffloadEntriesInfoManager.initializeDeviceGlobalVarEntryInfo(
4999           /*MangledName=*/GetMDString(1),
5000           static_cast<OffloadEntriesInfoManager::OMPTargetGlobalVarEntryKind>(
5001               /*Flags=*/GetMDInt(2)),
5002           /*Order=*/GetMDInt(3));
5003       break;
5004     }
5005   }
5006 }
5007 
5008 bool OffloadEntriesInfoManager::empty() const {
5009   return OffloadEntriesTargetRegion.empty() &&
5010          OffloadEntriesDeviceGlobalVar.empty();
5011 }
5012 
5013 unsigned OffloadEntriesInfoManager::getTargetRegionEntryInfoCount(
5014     const TargetRegionEntryInfo &EntryInfo) const {
5015   auto It = OffloadEntriesTargetRegionCount.find(
5016       getTargetRegionEntryCountKey(EntryInfo));
5017   if (It == OffloadEntriesTargetRegionCount.end())
5018     return 0;
5019   return It->second;
5020 }
5021 
5022 void OffloadEntriesInfoManager::incrementTargetRegionEntryInfoCount(
5023     const TargetRegionEntryInfo &EntryInfo) {
5024   OffloadEntriesTargetRegionCount[getTargetRegionEntryCountKey(EntryInfo)] =
5025       EntryInfo.Count + 1;
5026 }
5027 
5028 /// Initialize target region entry.
5029 void OffloadEntriesInfoManager::initializeTargetRegionEntryInfo(
5030     const TargetRegionEntryInfo &EntryInfo, unsigned Order) {
5031   OffloadEntriesTargetRegion[EntryInfo] =
5032       OffloadEntryInfoTargetRegion(Order, /*Addr=*/nullptr, /*ID=*/nullptr,
5033                                    OMPTargetRegionEntryTargetRegion);
5034   ++OffloadingEntriesNum;
5035 }
5036 
5037 void OffloadEntriesInfoManager::registerTargetRegionEntryInfo(
5038     TargetRegionEntryInfo EntryInfo, Constant *Addr, Constant *ID,
5039     OMPTargetRegionEntryKind Flags) {
5040   assert(EntryInfo.Count == 0 && "expected default EntryInfo");
5041 
5042   // Update the EntryInfo with the next available count for this location.
5043   EntryInfo.Count = getTargetRegionEntryInfoCount(EntryInfo);
5044 
5045   // If we are emitting code for a target, the entry is already initialized,
5046   // only has to be registered.
5047   if (Config.isEmbedded()) {
5048     // This could happen if the device compilation is invoked standalone.
5049     if (!hasTargetRegionEntryInfo(EntryInfo)) {
5050       return;
5051     }
5052     auto &Entry = OffloadEntriesTargetRegion[EntryInfo];
5053     Entry.setAddress(Addr);
5054     Entry.setID(ID);
5055     Entry.setFlags(Flags);
5056   } else {
5057     if (Flags == OffloadEntriesInfoManager::OMPTargetRegionEntryTargetRegion &&
5058         hasTargetRegionEntryInfo(EntryInfo, /*IgnoreAddressId*/ true))
5059       return;
5060     assert(!hasTargetRegionEntryInfo(EntryInfo) &&
5061            "Target region entry already registered!");
5062     OffloadEntryInfoTargetRegion Entry(OffloadingEntriesNum, Addr, ID, Flags);
5063     OffloadEntriesTargetRegion[EntryInfo] = Entry;
5064     ++OffloadingEntriesNum;
5065   }
5066   incrementTargetRegionEntryInfoCount(EntryInfo);
5067 }
5068 
5069 bool OffloadEntriesInfoManager::hasTargetRegionEntryInfo(
5070     TargetRegionEntryInfo EntryInfo, bool IgnoreAddressId) const {
5071 
5072   // Update the EntryInfo with the next available count for this location.
5073   EntryInfo.Count = getTargetRegionEntryInfoCount(EntryInfo);
5074 
5075   auto It = OffloadEntriesTargetRegion.find(EntryInfo);
5076   if (It == OffloadEntriesTargetRegion.end()) {
5077     return false;
5078   }
5079   // Fail if this entry is already registered.
5080   if (!IgnoreAddressId && (It->second.getAddress() || It->second.getID()))
5081     return false;
5082   return true;
5083 }
5084 
5085 void OffloadEntriesInfoManager::actOnTargetRegionEntriesInfo(
5086     const OffloadTargetRegionEntryInfoActTy &Action) {
5087   // Scan all target region entries and perform the provided action.
5088   for (const auto &It : OffloadEntriesTargetRegion) {
5089     Action(It.first, It.second);
5090   }
5091 }
5092 
5093 void OffloadEntriesInfoManager::initializeDeviceGlobalVarEntryInfo(
5094     StringRef Name, OMPTargetGlobalVarEntryKind Flags, unsigned Order) {
5095   OffloadEntriesDeviceGlobalVar.try_emplace(Name, Order, Flags);
5096   ++OffloadingEntriesNum;
5097 }
5098 
5099 void OffloadEntriesInfoManager::registerDeviceGlobalVarEntryInfo(
5100     StringRef VarName, Constant *Addr, int64_t VarSize,
5101     OMPTargetGlobalVarEntryKind Flags, GlobalValue::LinkageTypes Linkage) {
5102   if (Config.isEmbedded()) {
5103     // This could happen if the device compilation is invoked standalone.
5104     if (!hasDeviceGlobalVarEntryInfo(VarName))
5105       return;
5106     auto &Entry = OffloadEntriesDeviceGlobalVar[VarName];
5107     if (Entry.getAddress() && hasDeviceGlobalVarEntryInfo(VarName)) {
5108       if (Entry.getVarSize() == 0) {
5109         Entry.setVarSize(VarSize);
5110         Entry.setLinkage(Linkage);
5111       }
5112       return;
5113     }
5114     Entry.setVarSize(VarSize);
5115     Entry.setLinkage(Linkage);
5116     Entry.setAddress(Addr);
5117   } else {
5118     if (hasDeviceGlobalVarEntryInfo(VarName)) {
5119       auto &Entry = OffloadEntriesDeviceGlobalVar[VarName];
5120       assert(Entry.isValid() && Entry.getFlags() == Flags &&
5121              "Entry not initialized!");
5122       if (Entry.getVarSize() == 0) {
5123         Entry.setVarSize(VarSize);
5124         Entry.setLinkage(Linkage);
5125       }
5126       return;
5127     }
5128     OffloadEntriesDeviceGlobalVar.try_emplace(VarName, OffloadingEntriesNum,
5129                                               Addr, VarSize, Flags, Linkage);
5130     ++OffloadingEntriesNum;
5131   }
5132 }
5133 
5134 void OffloadEntriesInfoManager::actOnDeviceGlobalVarEntriesInfo(
5135     const OffloadDeviceGlobalVarEntryInfoActTy &Action) {
5136   // Scan all target region entries and perform the provided action.
5137   for (const auto &E : OffloadEntriesDeviceGlobalVar)
5138     Action(E.getKey(), E.getValue());
5139 }
5140 
5141 void CanonicalLoopInfo::collectControlBlocks(
5142     SmallVectorImpl<BasicBlock *> &BBs) {
5143   // We only count those BBs as control block for which we do not need to
5144   // reverse the CFG, i.e. not the loop body which can contain arbitrary control
5145   // flow. For consistency, this also means we do not add the Body block, which
5146   // is just the entry to the body code.
5147   BBs.reserve(BBs.size() + 6);
5148   BBs.append({getPreheader(), Header, Cond, Latch, Exit, getAfter()});
5149 }
5150 
5151 BasicBlock *CanonicalLoopInfo::getPreheader() const {
5152   assert(isValid() && "Requires a valid canonical loop");
5153   for (BasicBlock *Pred : predecessors(Header)) {
5154     if (Pred != Latch)
5155       return Pred;
5156   }
5157   llvm_unreachable("Missing preheader");
5158 }
5159 
5160 void CanonicalLoopInfo::setTripCount(Value *TripCount) {
5161   assert(isValid() && "Requires a valid canonical loop");
5162 
5163   Instruction *CmpI = &getCond()->front();
5164   assert(isa<CmpInst>(CmpI) && "First inst must compare IV with TripCount");
5165   CmpI->setOperand(1, TripCount);
5166 
5167 #ifndef NDEBUG
5168   assertOK();
5169 #endif
5170 }
5171 
5172 void CanonicalLoopInfo::mapIndVar(
5173     llvm::function_ref<Value *(Instruction *)> Updater) {
5174   assert(isValid() && "Requires a valid canonical loop");
5175 
5176   Instruction *OldIV = getIndVar();
5177 
5178   // Record all uses excluding those introduced by the updater. Uses by the
5179   // CanonicalLoopInfo itself to keep track of the number of iterations are
5180   // excluded.
5181   SmallVector<Use *> ReplacableUses;
5182   for (Use &U : OldIV->uses()) {
5183     auto *User = dyn_cast<Instruction>(U.getUser());
5184     if (!User)
5185       continue;
5186     if (User->getParent() == getCond())
5187       continue;
5188     if (User->getParent() == getLatch())
5189       continue;
5190     ReplacableUses.push_back(&U);
5191   }
5192 
5193   // Run the updater that may introduce new uses
5194   Value *NewIV = Updater(OldIV);
5195 
5196   // Replace the old uses with the value returned by the updater.
5197   for (Use *U : ReplacableUses)
5198     U->set(NewIV);
5199 
5200 #ifndef NDEBUG
5201   assertOK();
5202 #endif
5203 }
5204 
5205 void CanonicalLoopInfo::assertOK() const {
5206 #ifndef NDEBUG
5207   // No constraints if this object currently does not describe a loop.
5208   if (!isValid())
5209     return;
5210 
5211   BasicBlock *Preheader = getPreheader();
5212   BasicBlock *Body = getBody();
5213   BasicBlock *After = getAfter();
5214 
5215   // Verify standard control-flow we use for OpenMP loops.
5216   assert(Preheader);
5217   assert(isa<BranchInst>(Preheader->getTerminator()) &&
5218          "Preheader must terminate with unconditional branch");
5219   assert(Preheader->getSingleSuccessor() == Header &&
5220          "Preheader must jump to header");
5221 
5222   assert(Header);
5223   assert(isa<BranchInst>(Header->getTerminator()) &&
5224          "Header must terminate with unconditional branch");
5225   assert(Header->getSingleSuccessor() == Cond &&
5226          "Header must jump to exiting block");
5227 
5228   assert(Cond);
5229   assert(Cond->getSinglePredecessor() == Header &&
5230          "Exiting block only reachable from header");
5231 
5232   assert(isa<BranchInst>(Cond->getTerminator()) &&
5233          "Exiting block must terminate with conditional branch");
5234   assert(size(successors(Cond)) == 2 &&
5235          "Exiting block must have two successors");
5236   assert(cast<BranchInst>(Cond->getTerminator())->getSuccessor(0) == Body &&
5237          "Exiting block's first successor jump to the body");
5238   assert(cast<BranchInst>(Cond->getTerminator())->getSuccessor(1) == Exit &&
5239          "Exiting block's second successor must exit the loop");
5240 
5241   assert(Body);
5242   assert(Body->getSinglePredecessor() == Cond &&
5243          "Body only reachable from exiting block");
5244   assert(!isa<PHINode>(Body->front()));
5245 
5246   assert(Latch);
5247   assert(isa<BranchInst>(Latch->getTerminator()) &&
5248          "Latch must terminate with unconditional branch");
5249   assert(Latch->getSingleSuccessor() == Header && "Latch must jump to header");
5250   // TODO: To support simple redirecting of the end of the body code that has
5251   // multiple; introduce another auxiliary basic block like preheader and after.
5252   assert(Latch->getSinglePredecessor() != nullptr);
5253   assert(!isa<PHINode>(Latch->front()));
5254 
5255   assert(Exit);
5256   assert(isa<BranchInst>(Exit->getTerminator()) &&
5257          "Exit block must terminate with unconditional branch");
5258   assert(Exit->getSingleSuccessor() == After &&
5259          "Exit block must jump to after block");
5260 
5261   assert(After);
5262   assert(After->getSinglePredecessor() == Exit &&
5263          "After block only reachable from exit block");
5264   assert(After->empty() || !isa<PHINode>(After->front()));
5265 
5266   Instruction *IndVar = getIndVar();
5267   assert(IndVar && "Canonical induction variable not found?");
5268   assert(isa<IntegerType>(IndVar->getType()) &&
5269          "Induction variable must be an integer");
5270   assert(cast<PHINode>(IndVar)->getParent() == Header &&
5271          "Induction variable must be a PHI in the loop header");
5272   assert(cast<PHINode>(IndVar)->getIncomingBlock(0) == Preheader);
5273   assert(
5274       cast<ConstantInt>(cast<PHINode>(IndVar)->getIncomingValue(0))->isZero());
5275   assert(cast<PHINode>(IndVar)->getIncomingBlock(1) == Latch);
5276 
5277   auto *NextIndVar = cast<PHINode>(IndVar)->getIncomingValue(1);
5278   assert(cast<Instruction>(NextIndVar)->getParent() == Latch);
5279   assert(cast<BinaryOperator>(NextIndVar)->getOpcode() == BinaryOperator::Add);
5280   assert(cast<BinaryOperator>(NextIndVar)->getOperand(0) == IndVar);
5281   assert(cast<ConstantInt>(cast<BinaryOperator>(NextIndVar)->getOperand(1))
5282              ->isOne());
5283 
5284   Value *TripCount = getTripCount();
5285   assert(TripCount && "Loop trip count not found?");
5286   assert(IndVar->getType() == TripCount->getType() &&
5287          "Trip count and induction variable must have the same type");
5288 
5289   auto *CmpI = cast<CmpInst>(&Cond->front());
5290   assert(CmpI->getPredicate() == CmpInst::ICMP_ULT &&
5291          "Exit condition must be a signed less-than comparison");
5292   assert(CmpI->getOperand(0) == IndVar &&
5293          "Exit condition must compare the induction variable");
5294   assert(CmpI->getOperand(1) == TripCount &&
5295          "Exit condition must compare with the trip count");
5296 #endif
5297 }
5298 
5299 void CanonicalLoopInfo::invalidate() {
5300   Header = nullptr;
5301   Cond = nullptr;
5302   Latch = nullptr;
5303   Exit = nullptr;
5304 }
5305