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