1 //===-- IPO/OpenMPOpt.cpp - Collection of OpenMP specific optimizations ---===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // OpenMP specific optimizations:
10 //
11 // - Deduplication of runtime calls, e.g., omp_get_thread_num.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/Transforms/IPO/OpenMPOpt.h"
16 
17 #include "llvm/ADT/EnumeratedArray.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/CallGraph.h"
20 #include "llvm/Analysis/CallGraphSCCPass.h"
21 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
22 #include "llvm/Frontend/OpenMP/OMPConstants.h"
23 #include "llvm/Frontend/OpenMP/OMPIRBuilder.h"
24 #include "llvm/InitializePasses.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Transforms/IPO.h"
27 #include "llvm/Transforms/IPO/Attributor.h"
28 #include "llvm/Transforms/Utils/CallGraphUpdater.h"
29 
30 using namespace llvm;
31 using namespace omp;
32 
33 #define DEBUG_TYPE "openmp-opt"
34 
35 static cl::opt<bool> DisableOpenMPOptimizations(
36     "openmp-opt-disable", cl::ZeroOrMore,
37     cl::desc("Disable OpenMP specific optimizations."), cl::Hidden,
38     cl::init(false));
39 
40 static cl::opt<bool> PrintICVValues("openmp-print-icv-values", cl::init(false),
41                                     cl::Hidden);
42 static cl::opt<bool> PrintOpenMPKernels("openmp-print-gpu-kernels",
43                                         cl::init(false), cl::Hidden);
44 
45 STATISTIC(NumOpenMPRuntimeCallsDeduplicated,
46           "Number of OpenMP runtime calls deduplicated");
47 STATISTIC(NumOpenMPParallelRegionsDeleted,
48           "Number of OpenMP parallel regions deleted");
49 STATISTIC(NumOpenMPRuntimeFunctionsIdentified,
50           "Number of OpenMP runtime functions identified");
51 STATISTIC(NumOpenMPRuntimeFunctionUsesIdentified,
52           "Number of OpenMP runtime function uses identified");
53 STATISTIC(NumOpenMPTargetRegionKernels,
54           "Number of OpenMP target region entry points (=kernels) identified");
55 STATISTIC(
56     NumOpenMPParallelRegionsReplacedInGPUStateMachine,
57     "Number of OpenMP parallel regions replaced with ID in GPU state machines");
58 
59 #if !defined(NDEBUG)
60 static constexpr auto TAG = "[" DEBUG_TYPE "]";
61 #endif
62 
63 /// Apply \p CB to all uses of \p F. If \p LookThroughConstantExprUses is
64 /// true, constant expression users are not given to \p CB but their uses are
65 /// traversed transitively.
66 template <typename CBTy>
67 static void foreachUse(Function &F, CBTy CB,
68                        bool LookThroughConstantExprUses = true) {
69   SmallVector<Use *, 8> Worklist(make_pointer_range(F.uses()));
70 
71   for (unsigned idx = 0; idx < Worklist.size(); ++idx) {
72     Use &U = *Worklist[idx];
73 
74     // Allow use in constant bitcasts and simply look through them.
75     if (LookThroughConstantExprUses && isa<ConstantExpr>(U.getUser())) {
76       for (Use &CEU : cast<ConstantExpr>(U.getUser())->uses())
77         Worklist.push_back(&CEU);
78       continue;
79     }
80 
81     CB(U);
82   }
83 }
84 
85 /// Helper struct to store tracked ICV values at specif instructions.
86 struct ICVValue {
87   Instruction *Inst;
88   Value *TrackedValue;
89 
90   ICVValue(Instruction *I, Value *Val) : Inst(I), TrackedValue(Val) {}
91 };
92 
93 namespace llvm {
94 
95 // Provide DenseMapInfo for ICVValue
96 template <> struct DenseMapInfo<ICVValue> {
97   using InstInfo = DenseMapInfo<Instruction *>;
98   using ValueInfo = DenseMapInfo<Value *>;
99 
100   static inline ICVValue getEmptyKey() {
101     return ICVValue(InstInfo::getEmptyKey(), ValueInfo::getEmptyKey());
102   };
103 
104   static inline ICVValue getTombstoneKey() {
105     return ICVValue(InstInfo::getTombstoneKey(), ValueInfo::getTombstoneKey());
106   };
107 
108   static unsigned getHashValue(const ICVValue &ICVVal) {
109     return detail::combineHashValue(
110         InstInfo::getHashValue(ICVVal.Inst),
111         ValueInfo::getHashValue(ICVVal.TrackedValue));
112   }
113 
114   static bool isEqual(const ICVValue &LHS, const ICVValue &RHS) {
115     return InstInfo::isEqual(LHS.Inst, RHS.Inst) &&
116            ValueInfo::isEqual(LHS.TrackedValue, RHS.TrackedValue);
117   }
118 };
119 
120 } // end namespace llvm
121 
122 namespace {
123 
124 struct AAICVTracker;
125 
126 /// OpenMP specific information. For now, stores RFIs and ICVs also needed for
127 /// Attributor runs.
128 struct OMPInformationCache : public InformationCache {
129   OMPInformationCache(Module &M, AnalysisGetter &AG,
130                       BumpPtrAllocator &Allocator, SetVector<Function *> &CGSCC,
131                       SmallPtrSetImpl<Kernel> &Kernels)
132       : InformationCache(M, AG, Allocator, &CGSCC), OMPBuilder(M),
133         Kernels(Kernels) {
134     initializeModuleSlice(CGSCC);
135 
136     OMPBuilder.initialize();
137     initializeRuntimeFunctions();
138     initializeInternalControlVars();
139   }
140 
141   /// Generic information that describes an internal control variable.
142   struct InternalControlVarInfo {
143     /// The kind, as described by InternalControlVar enum.
144     InternalControlVar Kind;
145 
146     /// The name of the ICV.
147     StringRef Name;
148 
149     /// Environment variable associated with this ICV.
150     StringRef EnvVarName;
151 
152     /// Initial value kind.
153     ICVInitValue InitKind;
154 
155     /// Initial value.
156     ConstantInt *InitValue;
157 
158     /// Setter RTL function associated with this ICV.
159     RuntimeFunction Setter;
160 
161     /// Getter RTL function associated with this ICV.
162     RuntimeFunction Getter;
163 
164     /// RTL Function corresponding to the override clause of this ICV
165     RuntimeFunction Clause;
166   };
167 
168   /// Generic information that describes a runtime function
169   struct RuntimeFunctionInfo {
170 
171     /// The kind, as described by the RuntimeFunction enum.
172     RuntimeFunction Kind;
173 
174     /// The name of the function.
175     StringRef Name;
176 
177     /// Flag to indicate a variadic function.
178     bool IsVarArg;
179 
180     /// The return type of the function.
181     Type *ReturnType;
182 
183     /// The argument types of the function.
184     SmallVector<Type *, 8> ArgumentTypes;
185 
186     /// The declaration if available.
187     Function *Declaration = nullptr;
188 
189     /// Uses of this runtime function per function containing the use.
190     using UseVector = SmallVector<Use *, 16>;
191 
192     /// Clear UsesMap for runtime function.
193     void clearUsesMap() { UsesMap.clear(); }
194 
195     /// Boolean conversion that is true if the runtime function was found.
196     operator bool() const { return Declaration; }
197 
198     /// Return the vector of uses in function \p F.
199     UseVector &getOrCreateUseVector(Function *F) {
200       std::shared_ptr<UseVector> &UV = UsesMap[F];
201       if (!UV)
202         UV = std::make_shared<UseVector>();
203       return *UV;
204     }
205 
206     /// Return the vector of uses in function \p F or `nullptr` if there are
207     /// none.
208     const UseVector *getUseVector(Function &F) const {
209       auto I = UsesMap.find(&F);
210       if (I != UsesMap.end())
211         return I->second.get();
212       return nullptr;
213     }
214 
215     /// Return how many functions contain uses of this runtime function.
216     size_t getNumFunctionsWithUses() const { return UsesMap.size(); }
217 
218     /// Return the number of arguments (or the minimal number for variadic
219     /// functions).
220     size_t getNumArgs() const { return ArgumentTypes.size(); }
221 
222     /// Run the callback \p CB on each use and forget the use if the result is
223     /// true. The callback will be fed the function in which the use was
224     /// encountered as second argument.
225     void foreachUse(SmallVectorImpl<Function *> &SCC,
226                     function_ref<bool(Use &, Function &)> CB) {
227       for (Function *F : SCC)
228         foreachUse(CB, F);
229     }
230 
231     /// Run the callback \p CB on each use within the function \p F and forget
232     /// the use if the result is true.
233     void foreachUse(function_ref<bool(Use &, Function &)> CB, Function *F) {
234       SmallVector<unsigned, 8> ToBeDeleted;
235       ToBeDeleted.clear();
236 
237       unsigned Idx = 0;
238       UseVector &UV = getOrCreateUseVector(F);
239 
240       for (Use *U : UV) {
241         if (CB(*U, *F))
242           ToBeDeleted.push_back(Idx);
243         ++Idx;
244       }
245 
246       // Remove the to-be-deleted indices in reverse order as prior
247       // modifications will not modify the smaller indices.
248       while (!ToBeDeleted.empty()) {
249         unsigned Idx = ToBeDeleted.pop_back_val();
250         UV[Idx] = UV.back();
251         UV.pop_back();
252       }
253     }
254 
255   private:
256     /// Map from functions to all uses of this runtime function contained in
257     /// them.
258     DenseMap<Function *, std::shared_ptr<UseVector>> UsesMap;
259   };
260 
261   /// Initialize the ModuleSlice member based on \p SCC. ModuleSlices contains
262   /// (a subset of) all functions that we can look at during this SCC traversal.
263   /// This includes functions (transitively) called from the SCC and the
264   /// (transitive) callers of SCC functions. We also can look at a function if
265   /// there is a "reference edge", i.a., if the function somehow uses (!=calls)
266   /// a function in the SCC or a caller of a function in the SCC.
267   void initializeModuleSlice(SetVector<Function *> &SCC) {
268     ModuleSlice.insert(SCC.begin(), SCC.end());
269 
270     SmallPtrSet<Function *, 16> Seen;
271     SmallVector<Function *, 16> Worklist(SCC.begin(), SCC.end());
272     while (!Worklist.empty()) {
273       Function *F = Worklist.pop_back_val();
274       ModuleSlice.insert(F);
275 
276       for (Instruction &I : instructions(*F))
277         if (auto *CB = dyn_cast<CallBase>(&I))
278           if (Function *Callee = CB->getCalledFunction())
279             if (Seen.insert(Callee).second)
280               Worklist.push_back(Callee);
281     }
282 
283     Seen.clear();
284     Worklist.append(SCC.begin(), SCC.end());
285     while (!Worklist.empty()) {
286       Function *F = Worklist.pop_back_val();
287       ModuleSlice.insert(F);
288 
289       // Traverse all transitive uses.
290       foreachUse(*F, [&](Use &U) {
291         if (auto *UsrI = dyn_cast<Instruction>(U.getUser()))
292           if (Seen.insert(UsrI->getFunction()).second)
293             Worklist.push_back(UsrI->getFunction());
294       });
295     }
296   }
297 
298   /// The slice of the module we are allowed to look at.
299   SmallPtrSet<Function *, 8> ModuleSlice;
300 
301   /// An OpenMP-IR-Builder instance
302   OpenMPIRBuilder OMPBuilder;
303 
304   /// Map from runtime function kind to the runtime function description.
305   EnumeratedArray<RuntimeFunctionInfo, RuntimeFunction,
306                   RuntimeFunction::OMPRTL___last>
307       RFIs;
308 
309   /// Map from ICV kind to the ICV description.
310   EnumeratedArray<InternalControlVarInfo, InternalControlVar,
311                   InternalControlVar::ICV___last>
312       ICVs;
313 
314   /// Helper to initialize all internal control variable information for those
315   /// defined in OMPKinds.def.
316   void initializeInternalControlVars() {
317 #define ICV_RT_SET(_Name, RTL)                                                 \
318   {                                                                            \
319     auto &ICV = ICVs[_Name];                                                   \
320     ICV.Setter = RTL;                                                          \
321   }
322 #define ICV_RT_GET(Name, RTL)                                                  \
323   {                                                                            \
324     auto &ICV = ICVs[Name];                                                    \
325     ICV.Getter = RTL;                                                          \
326   }
327 #define ICV_DATA_ENV(Enum, _Name, _EnvVarName, Init)                           \
328   {                                                                            \
329     auto &ICV = ICVs[Enum];                                                    \
330     ICV.Name = _Name;                                                          \
331     ICV.Kind = Enum;                                                           \
332     ICV.InitKind = Init;                                                       \
333     ICV.EnvVarName = _EnvVarName;                                              \
334     switch (ICV.InitKind) {                                                    \
335     case ICV_IMPLEMENTATION_DEFINED:                                           \
336       ICV.InitValue = nullptr;                                                 \
337       break;                                                                   \
338     case ICV_ZERO:                                                             \
339       ICV.InitValue = ConstantInt::get(                                        \
340           Type::getInt32Ty(OMPBuilder.Int32->getContext()), 0);                \
341       break;                                                                   \
342     case ICV_FALSE:                                                            \
343       ICV.InitValue = ConstantInt::getFalse(OMPBuilder.Int1->getContext());    \
344       break;                                                                   \
345     case ICV_LAST:                                                             \
346       break;                                                                   \
347     }                                                                          \
348   }
349 #include "llvm/Frontend/OpenMP/OMPKinds.def"
350   }
351 
352   /// Returns true if the function declaration \p F matches the runtime
353   /// function types, that is, return type \p RTFRetType, and argument types
354   /// \p RTFArgTypes.
355   static bool declMatchesRTFTypes(Function *F, Type *RTFRetType,
356                                   SmallVector<Type *, 8> &RTFArgTypes) {
357     // TODO: We should output information to the user (under debug output
358     //       and via remarks).
359 
360     if (!F)
361       return false;
362     if (F->getReturnType() != RTFRetType)
363       return false;
364     if (F->arg_size() != RTFArgTypes.size())
365       return false;
366 
367     auto RTFTyIt = RTFArgTypes.begin();
368     for (Argument &Arg : F->args()) {
369       if (Arg.getType() != *RTFTyIt)
370         return false;
371 
372       ++RTFTyIt;
373     }
374 
375     return true;
376   }
377 
378   // Helper to collect all uses of the declaration in the UsesMap.
379   unsigned collectUses(RuntimeFunctionInfo &RFI, bool CollectStats = true) {
380     unsigned NumUses = 0;
381     if (!RFI.Declaration)
382       return NumUses;
383     OMPBuilder.addAttributes(RFI.Kind, *RFI.Declaration);
384 
385     if (CollectStats) {
386       NumOpenMPRuntimeFunctionsIdentified += 1;
387       NumOpenMPRuntimeFunctionUsesIdentified += RFI.Declaration->getNumUses();
388     }
389 
390     // TODO: We directly convert uses into proper calls and unknown uses.
391     for (Use &U : RFI.Declaration->uses()) {
392       if (Instruction *UserI = dyn_cast<Instruction>(U.getUser())) {
393         if (ModuleSlice.count(UserI->getFunction())) {
394           RFI.getOrCreateUseVector(UserI->getFunction()).push_back(&U);
395           ++NumUses;
396         }
397       } else {
398         RFI.getOrCreateUseVector(nullptr).push_back(&U);
399         ++NumUses;
400       }
401     }
402     return NumUses;
403   }
404 
405   // Helper function to recollect uses of all runtime functions.
406   void recollectUses() {
407     for (int Idx = 0; Idx < RFIs.size(); ++Idx) {
408       auto &RFI = RFIs[static_cast<RuntimeFunction>(Idx)];
409       RFI.clearUsesMap();
410       collectUses(RFI, /*CollectStats*/ false);
411     }
412   }
413 
414   /// Helper to initialize all runtime function information for those defined
415   /// in OpenMPKinds.def.
416   void initializeRuntimeFunctions() {
417     Module &M = *((*ModuleSlice.begin())->getParent());
418 
419     // Helper macros for handling __VA_ARGS__ in OMP_RTL
420 #define OMP_TYPE(VarName, ...)                                                 \
421   Type *VarName = OMPBuilder.VarName;                                          \
422   (void)VarName;
423 
424 #define OMP_ARRAY_TYPE(VarName, ...)                                           \
425   ArrayType *VarName##Ty = OMPBuilder.VarName##Ty;                             \
426   (void)VarName##Ty;                                                           \
427   PointerType *VarName##PtrTy = OMPBuilder.VarName##PtrTy;                     \
428   (void)VarName##PtrTy;
429 
430 #define OMP_FUNCTION_TYPE(VarName, ...)                                        \
431   FunctionType *VarName = OMPBuilder.VarName;                                  \
432   (void)VarName;                                                               \
433   PointerType *VarName##Ptr = OMPBuilder.VarName##Ptr;                         \
434   (void)VarName##Ptr;
435 
436 #define OMP_STRUCT_TYPE(VarName, ...)                                          \
437   StructType *VarName = OMPBuilder.VarName;                                    \
438   (void)VarName;                                                               \
439   PointerType *VarName##Ptr = OMPBuilder.VarName##Ptr;                         \
440   (void)VarName##Ptr;
441 
442 #define OMP_RTL(_Enum, _Name, _IsVarArg, _ReturnType, ...)                     \
443   {                                                                            \
444     SmallVector<Type *, 8> ArgsTypes({__VA_ARGS__});                           \
445     Function *F = M.getFunction(_Name);                                        \
446     if (declMatchesRTFTypes(F, OMPBuilder._ReturnType, ArgsTypes)) {           \
447       auto &RFI = RFIs[_Enum];                                                 \
448       RFI.Kind = _Enum;                                                        \
449       RFI.Name = _Name;                                                        \
450       RFI.IsVarArg = _IsVarArg;                                                \
451       RFI.ReturnType = OMPBuilder._ReturnType;                                 \
452       RFI.ArgumentTypes = std::move(ArgsTypes);                                \
453       RFI.Declaration = F;                                                     \
454       unsigned NumUses = collectUses(RFI);                                     \
455       (void)NumUses;                                                           \
456       LLVM_DEBUG({                                                             \
457         dbgs() << TAG << RFI.Name << (RFI.Declaration ? "" : " not")           \
458                << " found\n";                                                  \
459         if (RFI.Declaration)                                                   \
460           dbgs() << TAG << "-> got " << NumUses << " uses in "                 \
461                  << RFI.getNumFunctionsWithUses()                              \
462                  << " different functions.\n";                                 \
463       });                                                                      \
464     }                                                                          \
465   }
466 #include "llvm/Frontend/OpenMP/OMPKinds.def"
467 
468     // TODO: We should attach the attributes defined in OMPKinds.def.
469   }
470 
471   /// Collection of known kernels (\see Kernel) in the module.
472   SmallPtrSetImpl<Kernel> &Kernels;
473 };
474 
475 struct OpenMPOpt {
476 
477   using OptimizationRemarkGetter =
478       function_ref<OptimizationRemarkEmitter &(Function *)>;
479 
480   OpenMPOpt(SmallVectorImpl<Function *> &SCC, CallGraphUpdater &CGUpdater,
481             OptimizationRemarkGetter OREGetter,
482             OMPInformationCache &OMPInfoCache, Attributor &A)
483       : M(*(*SCC.begin())->getParent()), SCC(SCC), CGUpdater(CGUpdater),
484         OREGetter(OREGetter), OMPInfoCache(OMPInfoCache), A(A) {}
485 
486   /// Run all OpenMP optimizations on the underlying SCC/ModuleSlice.
487   bool run() {
488     if (SCC.empty())
489       return false;
490 
491     bool Changed = false;
492 
493     LLVM_DEBUG(dbgs() << TAG << "Run on SCC with " << SCC.size()
494                       << " functions in a slice with "
495                       << OMPInfoCache.ModuleSlice.size() << " functions\n");
496 
497     if (PrintICVValues)
498       printICVs();
499     if (PrintOpenMPKernels)
500       printKernels();
501 
502     Changed |= rewriteDeviceCodeStateMachine();
503 
504     Changed |= runAttributor();
505 
506     // Recollect uses, in case Attributor deleted any.
507     OMPInfoCache.recollectUses();
508 
509     Changed |= deduplicateRuntimeCalls();
510     Changed |= deleteParallelRegions();
511 
512     return Changed;
513   }
514 
515   /// Print initial ICV values for testing.
516   /// FIXME: This should be done from the Attributor once it is added.
517   void printICVs() const {
518     InternalControlVar ICVs[] = {ICV_nthreads, ICV_active_levels, ICV_cancel};
519 
520     for (Function *F : OMPInfoCache.ModuleSlice) {
521       for (auto ICV : ICVs) {
522         auto ICVInfo = OMPInfoCache.ICVs[ICV];
523         auto Remark = [&](OptimizationRemark OR) {
524           return OR << "OpenMP ICV " << ore::NV("OpenMPICV", ICVInfo.Name)
525                     << " Value: "
526                     << (ICVInfo.InitValue
527                             ? ICVInfo.InitValue->getValue().toString(10, true)
528                             : "IMPLEMENTATION_DEFINED");
529         };
530 
531         emitRemarkOnFunction(F, "OpenMPICVTracker", Remark);
532       }
533     }
534   }
535 
536   /// Print OpenMP GPU kernels for testing.
537   void printKernels() const {
538     for (Function *F : SCC) {
539       if (!OMPInfoCache.Kernels.count(F))
540         continue;
541 
542       auto Remark = [&](OptimizationRemark OR) {
543         return OR << "OpenMP GPU kernel "
544                   << ore::NV("OpenMPGPUKernel", F->getName()) << "\n";
545       };
546 
547       emitRemarkOnFunction(F, "OpenMPGPU", Remark);
548     }
549   }
550 
551   /// Return the call if \p U is a callee use in a regular call. If \p RFI is
552   /// given it has to be the callee or a nullptr is returned.
553   static CallInst *getCallIfRegularCall(
554       Use &U, OMPInformationCache::RuntimeFunctionInfo *RFI = nullptr) {
555     CallInst *CI = dyn_cast<CallInst>(U.getUser());
556     if (CI && CI->isCallee(&U) && !CI->hasOperandBundles() &&
557         (!RFI || CI->getCalledFunction() == RFI->Declaration))
558       return CI;
559     return nullptr;
560   }
561 
562   /// Return the call if \p V is a regular call. If \p RFI is given it has to be
563   /// the callee or a nullptr is returned.
564   static CallInst *getCallIfRegularCall(
565       Value &V, OMPInformationCache::RuntimeFunctionInfo *RFI = nullptr) {
566     CallInst *CI = dyn_cast<CallInst>(&V);
567     if (CI && !CI->hasOperandBundles() &&
568         (!RFI || CI->getCalledFunction() == RFI->Declaration))
569       return CI;
570     return nullptr;
571   }
572 
573 private:
574   /// Try to delete parallel regions if possible.
575   bool deleteParallelRegions() {
576     const unsigned CallbackCalleeOperand = 2;
577 
578     OMPInformationCache::RuntimeFunctionInfo &RFI =
579         OMPInfoCache.RFIs[OMPRTL___kmpc_fork_call];
580 
581     if (!RFI.Declaration)
582       return false;
583 
584     bool Changed = false;
585     auto DeleteCallCB = [&](Use &U, Function &) {
586       CallInst *CI = getCallIfRegularCall(U);
587       if (!CI)
588         return false;
589       auto *Fn = dyn_cast<Function>(
590           CI->getArgOperand(CallbackCalleeOperand)->stripPointerCasts());
591       if (!Fn)
592         return false;
593       if (!Fn->onlyReadsMemory())
594         return false;
595       if (!Fn->hasFnAttribute(Attribute::WillReturn))
596         return false;
597 
598       LLVM_DEBUG(dbgs() << TAG << "Delete read-only parallel region in "
599                         << CI->getCaller()->getName() << "\n");
600 
601       auto Remark = [&](OptimizationRemark OR) {
602         return OR << "Parallel region in "
603                   << ore::NV("OpenMPParallelDelete", CI->getCaller()->getName())
604                   << " deleted";
605       };
606       emitRemark<OptimizationRemark>(CI, "OpenMPParallelRegionDeletion",
607                                      Remark);
608 
609       CGUpdater.removeCallSite(*CI);
610       CI->eraseFromParent();
611       Changed = true;
612       ++NumOpenMPParallelRegionsDeleted;
613       return true;
614     };
615 
616     RFI.foreachUse(SCC, DeleteCallCB);
617 
618     return Changed;
619   }
620 
621   /// Try to eliminate runtime calls by reusing existing ones.
622   bool deduplicateRuntimeCalls() {
623     bool Changed = false;
624 
625     RuntimeFunction DeduplicableRuntimeCallIDs[] = {
626         OMPRTL_omp_get_num_threads,
627         OMPRTL_omp_in_parallel,
628         OMPRTL_omp_get_cancellation,
629         OMPRTL_omp_get_thread_limit,
630         OMPRTL_omp_get_supported_active_levels,
631         OMPRTL_omp_get_level,
632         OMPRTL_omp_get_ancestor_thread_num,
633         OMPRTL_omp_get_team_size,
634         OMPRTL_omp_get_active_level,
635         OMPRTL_omp_in_final,
636         OMPRTL_omp_get_proc_bind,
637         OMPRTL_omp_get_num_places,
638         OMPRTL_omp_get_num_procs,
639         OMPRTL_omp_get_place_num,
640         OMPRTL_omp_get_partition_num_places,
641         OMPRTL_omp_get_partition_place_nums};
642 
643     // Global-tid is handled separately.
644     SmallSetVector<Value *, 16> GTIdArgs;
645     collectGlobalThreadIdArguments(GTIdArgs);
646     LLVM_DEBUG(dbgs() << TAG << "Found " << GTIdArgs.size()
647                       << " global thread ID arguments\n");
648 
649     for (Function *F : SCC) {
650       for (auto DeduplicableRuntimeCallID : DeduplicableRuntimeCallIDs)
651         deduplicateRuntimeCalls(*F,
652                                 OMPInfoCache.RFIs[DeduplicableRuntimeCallID]);
653 
654       // __kmpc_global_thread_num is special as we can replace it with an
655       // argument in enough cases to make it worth trying.
656       Value *GTIdArg = nullptr;
657       for (Argument &Arg : F->args())
658         if (GTIdArgs.count(&Arg)) {
659           GTIdArg = &Arg;
660           break;
661         }
662       Changed |= deduplicateRuntimeCalls(
663           *F, OMPInfoCache.RFIs[OMPRTL___kmpc_global_thread_num], GTIdArg);
664     }
665 
666     return Changed;
667   }
668 
669   static Value *combinedIdentStruct(Value *CurrentIdent, Value *NextIdent,
670                                     bool GlobalOnly, bool &SingleChoice) {
671     if (CurrentIdent == NextIdent)
672       return CurrentIdent;
673 
674     // TODO: Figure out how to actually combine multiple debug locations. For
675     //       now we just keep an existing one if there is a single choice.
676     if (!GlobalOnly || isa<GlobalValue>(NextIdent)) {
677       SingleChoice = !CurrentIdent;
678       return NextIdent;
679     }
680     return nullptr;
681   }
682 
683   /// Return an `struct ident_t*` value that represents the ones used in the
684   /// calls of \p RFI inside of \p F. If \p GlobalOnly is true, we will not
685   /// return a local `struct ident_t*`. For now, if we cannot find a suitable
686   /// return value we create one from scratch. We also do not yet combine
687   /// information, e.g., the source locations, see combinedIdentStruct.
688   Value *
689   getCombinedIdentFromCallUsesIn(OMPInformationCache::RuntimeFunctionInfo &RFI,
690                                  Function &F, bool GlobalOnly) {
691     bool SingleChoice = true;
692     Value *Ident = nullptr;
693     auto CombineIdentStruct = [&](Use &U, Function &Caller) {
694       CallInst *CI = getCallIfRegularCall(U, &RFI);
695       if (!CI || &F != &Caller)
696         return false;
697       Ident = combinedIdentStruct(Ident, CI->getArgOperand(0),
698                                   /* GlobalOnly */ true, SingleChoice);
699       return false;
700     };
701     RFI.foreachUse(SCC, CombineIdentStruct);
702 
703     if (!Ident || !SingleChoice) {
704       // The IRBuilder uses the insertion block to get to the module, this is
705       // unfortunate but we work around it for now.
706       if (!OMPInfoCache.OMPBuilder.getInsertionPoint().getBlock())
707         OMPInfoCache.OMPBuilder.updateToLocation(OpenMPIRBuilder::InsertPointTy(
708             &F.getEntryBlock(), F.getEntryBlock().begin()));
709       // Create a fallback location if non was found.
710       // TODO: Use the debug locations of the calls instead.
711       Constant *Loc = OMPInfoCache.OMPBuilder.getOrCreateDefaultSrcLocStr();
712       Ident = OMPInfoCache.OMPBuilder.getOrCreateIdent(Loc);
713     }
714     return Ident;
715   }
716 
717   /// Try to eliminate calls of \p RFI in \p F by reusing an existing one or
718   /// \p ReplVal if given.
719   bool deduplicateRuntimeCalls(Function &F,
720                                OMPInformationCache::RuntimeFunctionInfo &RFI,
721                                Value *ReplVal = nullptr) {
722     auto *UV = RFI.getUseVector(F);
723     if (!UV || UV->size() + (ReplVal != nullptr) < 2)
724       return false;
725 
726     LLVM_DEBUG(
727         dbgs() << TAG << "Deduplicate " << UV->size() << " uses of " << RFI.Name
728                << (ReplVal ? " with an existing value\n" : "\n") << "\n");
729 
730     assert((!ReplVal || (isa<Argument>(ReplVal) &&
731                          cast<Argument>(ReplVal)->getParent() == &F)) &&
732            "Unexpected replacement value!");
733 
734     // TODO: Use dominance to find a good position instead.
735     auto CanBeMoved = [this](CallBase &CB) {
736       unsigned NumArgs = CB.getNumArgOperands();
737       if (NumArgs == 0)
738         return true;
739       if (CB.getArgOperand(0)->getType() != OMPInfoCache.OMPBuilder.IdentPtr)
740         return false;
741       for (unsigned u = 1; u < NumArgs; ++u)
742         if (isa<Instruction>(CB.getArgOperand(u)))
743           return false;
744       return true;
745     };
746 
747     if (!ReplVal) {
748       for (Use *U : *UV)
749         if (CallInst *CI = getCallIfRegularCall(*U, &RFI)) {
750           if (!CanBeMoved(*CI))
751             continue;
752 
753           auto Remark = [&](OptimizationRemark OR) {
754             auto newLoc = &*F.getEntryBlock().getFirstInsertionPt();
755             return OR << "OpenMP runtime call "
756                       << ore::NV("OpenMPOptRuntime", RFI.Name) << " moved to "
757                       << ore::NV("OpenMPRuntimeMoves", newLoc->getDebugLoc());
758           };
759           emitRemark<OptimizationRemark>(CI, "OpenMPRuntimeCodeMotion", Remark);
760 
761           CI->moveBefore(&*F.getEntryBlock().getFirstInsertionPt());
762           ReplVal = CI;
763           break;
764         }
765       if (!ReplVal)
766         return false;
767     }
768 
769     // If we use a call as a replacement value we need to make sure the ident is
770     // valid at the new location. For now we just pick a global one, either
771     // existing and used by one of the calls, or created from scratch.
772     if (CallBase *CI = dyn_cast<CallBase>(ReplVal)) {
773       if (CI->getNumArgOperands() > 0 &&
774           CI->getArgOperand(0)->getType() == OMPInfoCache.OMPBuilder.IdentPtr) {
775         Value *Ident = getCombinedIdentFromCallUsesIn(RFI, F,
776                                                       /* GlobalOnly */ true);
777         CI->setArgOperand(0, Ident);
778       }
779     }
780 
781     bool Changed = false;
782     auto ReplaceAndDeleteCB = [&](Use &U, Function &Caller) {
783       CallInst *CI = getCallIfRegularCall(U, &RFI);
784       if (!CI || CI == ReplVal || &F != &Caller)
785         return false;
786       assert(CI->getCaller() == &F && "Unexpected call!");
787 
788       auto Remark = [&](OptimizationRemark OR) {
789         return OR << "OpenMP runtime call "
790                   << ore::NV("OpenMPOptRuntime", RFI.Name) << " deduplicated";
791       };
792       emitRemark<OptimizationRemark>(CI, "OpenMPRuntimeDeduplicated", Remark);
793 
794       CGUpdater.removeCallSite(*CI);
795       CI->replaceAllUsesWith(ReplVal);
796       CI->eraseFromParent();
797       ++NumOpenMPRuntimeCallsDeduplicated;
798       Changed = true;
799       return true;
800     };
801     RFI.foreachUse(SCC, ReplaceAndDeleteCB);
802 
803     return Changed;
804   }
805 
806   /// Collect arguments that represent the global thread id in \p GTIdArgs.
807   void collectGlobalThreadIdArguments(SmallSetVector<Value *, 16> &GTIdArgs) {
808     // TODO: Below we basically perform a fixpoint iteration with a pessimistic
809     //       initialization. We could define an AbstractAttribute instead and
810     //       run the Attributor here once it can be run as an SCC pass.
811 
812     // Helper to check the argument \p ArgNo at all call sites of \p F for
813     // a GTId.
814     auto CallArgOpIsGTId = [&](Function &F, unsigned ArgNo, CallInst &RefCI) {
815       if (!F.hasLocalLinkage())
816         return false;
817       for (Use &U : F.uses()) {
818         if (CallInst *CI = getCallIfRegularCall(U)) {
819           Value *ArgOp = CI->getArgOperand(ArgNo);
820           if (CI == &RefCI || GTIdArgs.count(ArgOp) ||
821               getCallIfRegularCall(
822                   *ArgOp, &OMPInfoCache.RFIs[OMPRTL___kmpc_global_thread_num]))
823             continue;
824         }
825         return false;
826       }
827       return true;
828     };
829 
830     // Helper to identify uses of a GTId as GTId arguments.
831     auto AddUserArgs = [&](Value &GTId) {
832       for (Use &U : GTId.uses())
833         if (CallInst *CI = dyn_cast<CallInst>(U.getUser()))
834           if (CI->isArgOperand(&U))
835             if (Function *Callee = CI->getCalledFunction())
836               if (CallArgOpIsGTId(*Callee, U.getOperandNo(), *CI))
837                 GTIdArgs.insert(Callee->getArg(U.getOperandNo()));
838     };
839 
840     // The argument users of __kmpc_global_thread_num calls are GTIds.
841     OMPInformationCache::RuntimeFunctionInfo &GlobThreadNumRFI =
842         OMPInfoCache.RFIs[OMPRTL___kmpc_global_thread_num];
843 
844     GlobThreadNumRFI.foreachUse(SCC, [&](Use &U, Function &F) {
845       if (CallInst *CI = getCallIfRegularCall(U, &GlobThreadNumRFI))
846         AddUserArgs(*CI);
847       return false;
848     });
849 
850     // Transitively search for more arguments by looking at the users of the
851     // ones we know already. During the search the GTIdArgs vector is extended
852     // so we cannot cache the size nor can we use a range based for.
853     for (unsigned u = 0; u < GTIdArgs.size(); ++u)
854       AddUserArgs(*GTIdArgs[u]);
855   }
856 
857   /// Kernel (=GPU) optimizations and utility functions
858   ///
859   ///{{
860 
861   /// Check if \p F is a kernel, hence entry point for target offloading.
862   bool isKernel(Function &F) { return OMPInfoCache.Kernels.count(&F); }
863 
864   /// Cache to remember the unique kernel for a function.
865   DenseMap<Function *, Optional<Kernel>> UniqueKernelMap;
866 
867   /// Find the unique kernel that will execute \p F, if any.
868   Kernel getUniqueKernelFor(Function &F);
869 
870   /// Find the unique kernel that will execute \p I, if any.
871   Kernel getUniqueKernelFor(Instruction &I) {
872     return getUniqueKernelFor(*I.getFunction());
873   }
874 
875   /// Rewrite the device (=GPU) code state machine create in non-SPMD mode in
876   /// the cases we can avoid taking the address of a function.
877   bool rewriteDeviceCodeStateMachine();
878 
879   ///
880   ///}}
881 
882   /// Emit a remark generically
883   ///
884   /// This template function can be used to generically emit a remark. The
885   /// RemarkKind should be one of the following:
886   ///   - OptimizationRemark to indicate a successful optimization attempt
887   ///   - OptimizationRemarkMissed to report a failed optimization attempt
888   ///   - OptimizationRemarkAnalysis to provide additional information about an
889   ///     optimization attempt
890   ///
891   /// The remark is built using a callback function provided by the caller that
892   /// takes a RemarkKind as input and returns a RemarkKind.
893   template <typename RemarkKind,
894             typename RemarkCallBack = function_ref<RemarkKind(RemarkKind &&)>>
895   void emitRemark(Instruction *Inst, StringRef RemarkName,
896                   RemarkCallBack &&RemarkCB) const {
897     Function *F = Inst->getParent()->getParent();
898     auto &ORE = OREGetter(F);
899 
900     ORE.emit(
901         [&]() { return RemarkCB(RemarkKind(DEBUG_TYPE, RemarkName, Inst)); });
902   }
903 
904   /// Emit a remark on a function. Since only OptimizationRemark is supporting
905   /// this, it can't be made generic.
906   void
907   emitRemarkOnFunction(Function *F, StringRef RemarkName,
908                        function_ref<OptimizationRemark(OptimizationRemark &&)>
909                            &&RemarkCB) const {
910     auto &ORE = OREGetter(F);
911 
912     ORE.emit([&]() {
913       return RemarkCB(OptimizationRemark(DEBUG_TYPE, RemarkName, F));
914     });
915   }
916 
917   /// The underlying module.
918   Module &M;
919 
920   /// The SCC we are operating on.
921   SmallVectorImpl<Function *> &SCC;
922 
923   /// Callback to update the call graph, the first argument is a removed call,
924   /// the second an optional replacement call.
925   CallGraphUpdater &CGUpdater;
926 
927   /// Callback to get an OptimizationRemarkEmitter from a Function *
928   OptimizationRemarkGetter OREGetter;
929 
930   /// OpenMP-specific information cache. Also Used for Attributor runs.
931   OMPInformationCache &OMPInfoCache;
932 
933   /// Attributor instance.
934   Attributor &A;
935 
936   /// Helper function to run Attributor on SCC.
937   bool runAttributor() {
938     if (SCC.empty())
939       return false;
940 
941     registerAAs();
942 
943     ChangeStatus Changed = A.run();
944 
945     LLVM_DEBUG(dbgs() << "[Attributor] Done with " << SCC.size()
946                       << " functions, result: " << Changed << ".\n");
947 
948     return Changed == ChangeStatus::CHANGED;
949   }
950 
951   /// Populate the Attributor with abstract attribute opportunities in the
952   /// function.
953   void registerAAs() {
954     for (Function *F : SCC) {
955       if (F->isDeclaration())
956         continue;
957 
958       A.getOrCreateAAFor<AAICVTracker>(IRPosition::function(*F));
959     }
960   }
961 };
962 
963 Kernel OpenMPOpt::getUniqueKernelFor(Function &F) {
964   if (!OMPInfoCache.ModuleSlice.count(&F))
965     return nullptr;
966 
967   // Use a scope to keep the lifetime of the CachedKernel short.
968   {
969     Optional<Kernel> &CachedKernel = UniqueKernelMap[&F];
970     if (CachedKernel)
971       return *CachedKernel;
972 
973     // TODO: We should use an AA to create an (optimistic and callback
974     //       call-aware) call graph. For now we stick to simple patterns that
975     //       are less powerful, basically the worst fixpoint.
976     if (isKernel(F)) {
977       CachedKernel = Kernel(&F);
978       return *CachedKernel;
979     }
980 
981     CachedKernel = nullptr;
982     if (!F.hasLocalLinkage())
983       return nullptr;
984   }
985 
986   auto GetUniqueKernelForUse = [&](const Use &U) -> Kernel {
987     if (auto *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
988       // Allow use in equality comparisons.
989       if (Cmp->isEquality())
990         return getUniqueKernelFor(*Cmp);
991       return nullptr;
992     }
993     if (auto *CB = dyn_cast<CallBase>(U.getUser())) {
994       // Allow direct calls.
995       if (CB->isCallee(&U))
996         return getUniqueKernelFor(*CB);
997       // Allow the use in __kmpc_kernel_prepare_parallel calls.
998       if (Function *Callee = CB->getCalledFunction())
999         if (Callee->getName() == "__kmpc_kernel_prepare_parallel")
1000           return getUniqueKernelFor(*CB);
1001       return nullptr;
1002     }
1003     // Disallow every other use.
1004     return nullptr;
1005   };
1006 
1007   // TODO: In the future we want to track more than just a unique kernel.
1008   SmallPtrSet<Kernel, 2> PotentialKernels;
1009   foreachUse(F, [&](const Use &U) {
1010     PotentialKernels.insert(GetUniqueKernelForUse(U));
1011   });
1012 
1013   Kernel K = nullptr;
1014   if (PotentialKernels.size() == 1)
1015     K = *PotentialKernels.begin();
1016 
1017   // Cache the result.
1018   UniqueKernelMap[&F] = K;
1019 
1020   return K;
1021 }
1022 
1023 bool OpenMPOpt::rewriteDeviceCodeStateMachine() {
1024   OMPInformationCache::RuntimeFunctionInfo &KernelPrepareParallelRFI =
1025       OMPInfoCache.RFIs[OMPRTL___kmpc_kernel_prepare_parallel];
1026 
1027   bool Changed = false;
1028   if (!KernelPrepareParallelRFI)
1029     return Changed;
1030 
1031   for (Function *F : SCC) {
1032 
1033     // Check if the function is uses in a __kmpc_kernel_prepare_parallel call at
1034     // all.
1035     bool UnknownUse = false;
1036     bool KernelPrepareUse = false;
1037     unsigned NumDirectCalls = 0;
1038 
1039     SmallVector<Use *, 2> ToBeReplacedStateMachineUses;
1040     foreachUse(*F, [&](Use &U) {
1041       if (auto *CB = dyn_cast<CallBase>(U.getUser()))
1042         if (CB->isCallee(&U)) {
1043           ++NumDirectCalls;
1044           return;
1045         }
1046 
1047       if (isa<ICmpInst>(U.getUser())) {
1048         ToBeReplacedStateMachineUses.push_back(&U);
1049         return;
1050       }
1051       if (!KernelPrepareUse && OpenMPOpt::getCallIfRegularCall(
1052                                    *U.getUser(), &KernelPrepareParallelRFI)) {
1053         KernelPrepareUse = true;
1054         ToBeReplacedStateMachineUses.push_back(&U);
1055         return;
1056       }
1057       UnknownUse = true;
1058     });
1059 
1060     // Do not emit a remark if we haven't seen a __kmpc_kernel_prepare_parallel
1061     // use.
1062     if (!KernelPrepareUse)
1063       continue;
1064 
1065     {
1066       auto Remark = [&](OptimizationRemark OR) {
1067         return OR << "Found a parallel region that is called in a target "
1068                      "region but not part of a combined target construct nor "
1069                      "nesed inside a target construct without intermediate "
1070                      "code. This can lead to excessive register usage for "
1071                      "unrelated target regions in the same translation unit "
1072                      "due to spurious call edges assumed by ptxas.";
1073       };
1074       emitRemarkOnFunction(F, "OpenMPParallelRegionInNonSPMD", Remark);
1075     }
1076 
1077     // If this ever hits, we should investigate.
1078     // TODO: Checking the number of uses is not a necessary restriction and
1079     // should be lifted.
1080     if (UnknownUse || NumDirectCalls != 1 ||
1081         ToBeReplacedStateMachineUses.size() != 2) {
1082       {
1083         auto Remark = [&](OptimizationRemark OR) {
1084           return OR << "Parallel region is used in "
1085                     << (UnknownUse ? "unknown" : "unexpected")
1086                     << " ways; will not attempt to rewrite the state machine.";
1087         };
1088         emitRemarkOnFunction(F, "OpenMPParallelRegionInNonSPMD", Remark);
1089       }
1090       continue;
1091     }
1092 
1093     // Even if we have __kmpc_kernel_prepare_parallel calls, we (for now) give
1094     // up if the function is not called from a unique kernel.
1095     Kernel K = getUniqueKernelFor(*F);
1096     if (!K) {
1097       {
1098         auto Remark = [&](OptimizationRemark OR) {
1099           return OR << "Parallel region is not known to be called from a "
1100                        "unique single target region, maybe the surrounding "
1101                        "function has external linkage?; will not attempt to "
1102                        "rewrite the state machine use.";
1103         };
1104         emitRemarkOnFunction(F, "OpenMPParallelRegionInMultipleKernesl",
1105                              Remark);
1106       }
1107       continue;
1108     }
1109 
1110     // We now know F is a parallel body function called only from the kernel K.
1111     // We also identified the state machine uses in which we replace the
1112     // function pointer by a new global symbol for identification purposes. This
1113     // ensures only direct calls to the function are left.
1114 
1115     {
1116       auto RemarkParalleRegion = [&](OptimizationRemark OR) {
1117         return OR << "Specialize parallel region that is only reached from a "
1118                      "single target region to avoid spurious call edges and "
1119                      "excessive register usage in other target regions. "
1120                      "(parallel region ID: "
1121                   << ore::NV("OpenMPParallelRegion", F->getName())
1122                   << ", kernel ID: "
1123                   << ore::NV("OpenMPTargetRegion", K->getName()) << ")";
1124       };
1125       emitRemarkOnFunction(F, "OpenMPParallelRegionInNonSPMD",
1126                            RemarkParalleRegion);
1127       auto RemarkKernel = [&](OptimizationRemark OR) {
1128         return OR << "Target region containing the parallel region that is "
1129                      "specialized. (parallel region ID: "
1130                   << ore::NV("OpenMPParallelRegion", F->getName())
1131                   << ", kernel ID: "
1132                   << ore::NV("OpenMPTargetRegion", K->getName()) << ")";
1133       };
1134       emitRemarkOnFunction(K, "OpenMPParallelRegionInNonSPMD", RemarkKernel);
1135     }
1136 
1137     Module &M = *F->getParent();
1138     Type *Int8Ty = Type::getInt8Ty(M.getContext());
1139 
1140     auto *ID = new GlobalVariable(
1141         M, Int8Ty, /* isConstant */ true, GlobalValue::PrivateLinkage,
1142         UndefValue::get(Int8Ty), F->getName() + ".ID");
1143 
1144     for (Use *U : ToBeReplacedStateMachineUses)
1145       U->set(ConstantExpr::getBitCast(ID, U->get()->getType()));
1146 
1147     ++NumOpenMPParallelRegionsReplacedInGPUStateMachine;
1148 
1149     Changed = true;
1150   }
1151 
1152   return Changed;
1153 }
1154 
1155 /// Abstract Attribute for tracking ICV values.
1156 struct AAICVTracker : public StateWrapper<BooleanState, AbstractAttribute> {
1157   using Base = StateWrapper<BooleanState, AbstractAttribute>;
1158   AAICVTracker(const IRPosition &IRP, Attributor &A) : Base(IRP) {}
1159 
1160   /// Returns true if value is assumed to be tracked.
1161   bool isAssumedTracked() const { return getAssumed(); }
1162 
1163   /// Returns true if value is known to be tracked.
1164   bool isKnownTracked() const { return getAssumed(); }
1165 
1166   /// Create an abstract attribute biew for the position \p IRP.
1167   static AAICVTracker &createForPosition(const IRPosition &IRP, Attributor &A);
1168 
1169   /// Return the value with which \p I can be replaced for specific \p ICV.
1170   virtual Value *getReplacementValue(InternalControlVar ICV,
1171                                      const Instruction *I, Attributor &A) = 0;
1172 
1173   /// See AbstractAttribute::getName()
1174   const std::string getName() const override { return "AAICVTracker"; }
1175 
1176   /// See AbstractAttribute::getIdAddr()
1177   const char *getIdAddr() const override { return &ID; }
1178 
1179   /// This function should return true if the type of the \p AA is AAICVTracker
1180   static bool classof(const AbstractAttribute *AA) {
1181     return (AA->getIdAddr() == &ID);
1182   }
1183 
1184   static const char ID;
1185 };
1186 
1187 struct AAICVTrackerFunction : public AAICVTracker {
1188   AAICVTrackerFunction(const IRPosition &IRP, Attributor &A)
1189       : AAICVTracker(IRP, A) {}
1190 
1191   // FIXME: come up with better string.
1192   const std::string getAsStr() const override { return "ICVTracker"; }
1193 
1194   // FIXME: come up with some stats.
1195   void trackStatistics() const override {}
1196 
1197   /// TODO: decide whether to deduplicate here, or use current
1198   /// deduplicateRuntimeCalls function.
1199   ChangeStatus manifest(Attributor &A) override {
1200     ChangeStatus Changed = ChangeStatus::UNCHANGED;
1201 
1202     for (InternalControlVar &ICV : TrackableICVs)
1203       if (deduplicateICVGetters(ICV, A))
1204         Changed = ChangeStatus::CHANGED;
1205 
1206     return Changed;
1207   }
1208 
1209   bool deduplicateICVGetters(InternalControlVar &ICV, Attributor &A) {
1210     auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache());
1211     auto &ICVInfo = OMPInfoCache.ICVs[ICV];
1212     auto &GetterRFI = OMPInfoCache.RFIs[ICVInfo.Getter];
1213 
1214     bool Changed = false;
1215 
1216     auto ReplaceAndDeleteCB = [&](Use &U, Function &Caller) {
1217       CallInst *CI = OpenMPOpt::getCallIfRegularCall(U, &GetterRFI);
1218       Instruction *UserI = cast<Instruction>(U.getUser());
1219       Value *ReplVal = getReplacementValue(ICV, UserI, A);
1220 
1221       if (!ReplVal || !CI)
1222         return false;
1223 
1224       A.removeCallSite(CI);
1225       CI->replaceAllUsesWith(ReplVal);
1226       CI->eraseFromParent();
1227       Changed = true;
1228       return true;
1229     };
1230 
1231     GetterRFI.foreachUse(ReplaceAndDeleteCB, getAnchorScope());
1232     return Changed;
1233   }
1234 
1235   // Map of ICV to their values at specific program point.
1236   EnumeratedArray<SmallSetVector<ICVValue, 4>, InternalControlVar,
1237                   InternalControlVar::ICV___last>
1238       ICVValuesMap;
1239 
1240   // Currently only nthreads is being tracked.
1241   // this array will only grow with time.
1242   InternalControlVar TrackableICVs[1] = {ICV_nthreads};
1243 
1244   ChangeStatus updateImpl(Attributor &A) override {
1245     ChangeStatus HasChanged = ChangeStatus::UNCHANGED;
1246 
1247     Function *F = getAnchorScope();
1248 
1249     auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache());
1250 
1251     for (InternalControlVar ICV : TrackableICVs) {
1252       auto &SetterRFI = OMPInfoCache.RFIs[OMPInfoCache.ICVs[ICV].Setter];
1253 
1254       auto TrackValues = [&](Use &U, Function &) {
1255         CallInst *CI = OpenMPOpt::getCallIfRegularCall(U);
1256         if (!CI)
1257           return false;
1258 
1259         // FIXME: handle setters with more that 1 arguments.
1260         /// Track new value.
1261         if (ICVValuesMap[ICV].insert(ICVValue(CI, CI->getArgOperand(0))))
1262           HasChanged = ChangeStatus::CHANGED;
1263 
1264         return false;
1265       };
1266 
1267       SetterRFI.foreachUse(TrackValues, F);
1268     }
1269 
1270     return HasChanged;
1271   }
1272 
1273   /// Return the value with which \p I can be replaced for specific \p ICV.
1274   Value *getReplacementValue(InternalControlVar ICV, const Instruction *I,
1275                              Attributor &A) override {
1276     const BasicBlock *CurrBB = I->getParent();
1277 
1278     auto &ValuesSet = ICVValuesMap[ICV];
1279     auto &OMPInfoCache = static_cast<OMPInformationCache &>(A.getInfoCache());
1280     auto &GetterRFI = OMPInfoCache.RFIs[OMPInfoCache.ICVs[ICV].Getter];
1281 
1282     for (const auto &ICVVal : ValuesSet) {
1283       if (CurrBB == ICVVal.Inst->getParent()) {
1284         if (!ICVVal.Inst->comesBefore(I))
1285           continue;
1286 
1287         // both instructions are in the same BB and at \p I we know the ICV
1288         // value.
1289         while (I != ICVVal.Inst) {
1290           // we don't yet know if a call might update an ICV.
1291           // TODO: check callsite AA for value.
1292           if (const auto *CB = dyn_cast<CallBase>(I))
1293             if (CB->getCalledFunction() != GetterRFI.Declaration)
1294               return nullptr;
1295 
1296           I = I->getPrevNode();
1297         }
1298 
1299         // No call in between, return the value.
1300         return ICVVal.TrackedValue;
1301       }
1302     }
1303 
1304     // No value was tracked.
1305     return nullptr;
1306   }
1307 };
1308 } // namespace
1309 
1310 const char AAICVTracker::ID = 0;
1311 
1312 AAICVTracker &AAICVTracker::createForPosition(const IRPosition &IRP,
1313                                               Attributor &A) {
1314   AAICVTracker *AA = nullptr;
1315   switch (IRP.getPositionKind()) {
1316   case IRPosition::IRP_INVALID:
1317   case IRPosition::IRP_FLOAT:
1318   case IRPosition::IRP_ARGUMENT:
1319   case IRPosition::IRP_RETURNED:
1320   case IRPosition::IRP_CALL_SITE_RETURNED:
1321   case IRPosition::IRP_CALL_SITE_ARGUMENT:
1322   case IRPosition::IRP_CALL_SITE:
1323     llvm_unreachable("ICVTracker can only be created for function position!");
1324   case IRPosition::IRP_FUNCTION:
1325     AA = new (A.Allocator) AAICVTrackerFunction(IRP, A);
1326     break;
1327   }
1328 
1329   return *AA;
1330 }
1331 
1332 PreservedAnalyses OpenMPOptPass::run(LazyCallGraph::SCC &C,
1333                                      CGSCCAnalysisManager &AM,
1334                                      LazyCallGraph &CG, CGSCCUpdateResult &UR) {
1335   if (!containsOpenMP(*C.begin()->getFunction().getParent(), OMPInModule))
1336     return PreservedAnalyses::all();
1337 
1338   if (DisableOpenMPOptimizations)
1339     return PreservedAnalyses::all();
1340 
1341   SmallVector<Function *, 16> SCC;
1342   for (LazyCallGraph::Node &N : C)
1343     SCC.push_back(&N.getFunction());
1344 
1345   if (SCC.empty())
1346     return PreservedAnalyses::all();
1347 
1348   FunctionAnalysisManager &FAM =
1349       AM.getResult<FunctionAnalysisManagerCGSCCProxy>(C, CG).getManager();
1350 
1351   AnalysisGetter AG(FAM);
1352 
1353   auto OREGetter = [&FAM](Function *F) -> OptimizationRemarkEmitter & {
1354     return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F);
1355   };
1356 
1357   CallGraphUpdater CGUpdater;
1358   CGUpdater.initialize(CG, C, AM, UR);
1359 
1360   SetVector<Function *> Functions(SCC.begin(), SCC.end());
1361   BumpPtrAllocator Allocator;
1362   OMPInformationCache InfoCache(*(Functions.back()->getParent()), AG, Allocator,
1363                                 /*CGSCC*/ Functions, OMPInModule.getKernels());
1364 
1365   Attributor A(Functions, InfoCache, CGUpdater);
1366 
1367   // TODO: Compute the module slice we are allowed to look at.
1368   OpenMPOpt OMPOpt(SCC, CGUpdater, OREGetter, InfoCache, A);
1369   bool Changed = OMPOpt.run();
1370   if (Changed)
1371     return PreservedAnalyses::none();
1372 
1373   return PreservedAnalyses::all();
1374 }
1375 
1376 namespace {
1377 
1378 struct OpenMPOptLegacyPass : public CallGraphSCCPass {
1379   CallGraphUpdater CGUpdater;
1380   OpenMPInModule OMPInModule;
1381   static char ID;
1382 
1383   OpenMPOptLegacyPass() : CallGraphSCCPass(ID) {
1384     initializeOpenMPOptLegacyPassPass(*PassRegistry::getPassRegistry());
1385   }
1386 
1387   void getAnalysisUsage(AnalysisUsage &AU) const override {
1388     CallGraphSCCPass::getAnalysisUsage(AU);
1389   }
1390 
1391   bool doInitialization(CallGraph &CG) override {
1392     // Disable the pass if there is no OpenMP (runtime call) in the module.
1393     containsOpenMP(CG.getModule(), OMPInModule);
1394     return false;
1395   }
1396 
1397   bool runOnSCC(CallGraphSCC &CGSCC) override {
1398     if (!containsOpenMP(CGSCC.getCallGraph().getModule(), OMPInModule))
1399       return false;
1400     if (DisableOpenMPOptimizations || skipSCC(CGSCC))
1401       return false;
1402 
1403     SmallVector<Function *, 16> SCC;
1404     for (CallGraphNode *CGN : CGSCC)
1405       if (Function *Fn = CGN->getFunction())
1406         if (!Fn->isDeclaration())
1407           SCC.push_back(Fn);
1408 
1409     if (SCC.empty())
1410       return false;
1411 
1412     CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
1413     CGUpdater.initialize(CG, CGSCC);
1414 
1415     // Maintain a map of functions to avoid rebuilding the ORE
1416     DenseMap<Function *, std::unique_ptr<OptimizationRemarkEmitter>> OREMap;
1417     auto OREGetter = [&OREMap](Function *F) -> OptimizationRemarkEmitter & {
1418       std::unique_ptr<OptimizationRemarkEmitter> &ORE = OREMap[F];
1419       if (!ORE)
1420         ORE = std::make_unique<OptimizationRemarkEmitter>(F);
1421       return *ORE;
1422     };
1423 
1424     AnalysisGetter AG;
1425     SetVector<Function *> Functions(SCC.begin(), SCC.end());
1426     BumpPtrAllocator Allocator;
1427     OMPInformationCache InfoCache(
1428         *(Functions.back()->getParent()), AG, Allocator,
1429         /*CGSCC*/ Functions, OMPInModule.getKernels());
1430 
1431     Attributor A(Functions, InfoCache, CGUpdater);
1432 
1433     // TODO: Compute the module slice we are allowed to look at.
1434     OpenMPOpt OMPOpt(SCC, CGUpdater, OREGetter, InfoCache, A);
1435     return OMPOpt.run();
1436   }
1437 
1438   bool doFinalization(CallGraph &CG) override { return CGUpdater.finalize(); }
1439 };
1440 
1441 } // end anonymous namespace
1442 
1443 void OpenMPInModule::identifyKernels(Module &M) {
1444 
1445   NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations");
1446   if (!MD)
1447     return;
1448 
1449   for (auto *Op : MD->operands()) {
1450     if (Op->getNumOperands() < 2)
1451       continue;
1452     MDString *KindID = dyn_cast<MDString>(Op->getOperand(1));
1453     if (!KindID || KindID->getString() != "kernel")
1454       continue;
1455 
1456     Function *KernelFn =
1457         mdconst::dyn_extract_or_null<Function>(Op->getOperand(0));
1458     if (!KernelFn)
1459       continue;
1460 
1461     ++NumOpenMPTargetRegionKernels;
1462 
1463     Kernels.insert(KernelFn);
1464   }
1465 }
1466 
1467 bool llvm::omp::containsOpenMP(Module &M, OpenMPInModule &OMPInModule) {
1468   if (OMPInModule.isKnown())
1469     return OMPInModule;
1470 
1471   // MSVC doesn't like long if-else chains for some reason and instead just
1472   // issues an error. Work around it..
1473   do {
1474 #define OMP_RTL(_Enum, _Name, ...)                                             \
1475   if (M.getFunction(_Name)) {                                                  \
1476     OMPInModule = true;                                                        \
1477     break;                                                                     \
1478   }
1479 #include "llvm/Frontend/OpenMP/OMPKinds.def"
1480   } while (false);
1481 
1482   // Identify kernels once. TODO: We should split the OMPInformationCache into a
1483   // module and an SCC part. The kernel information, among other things, could
1484   // go into the module part.
1485   if (OMPInModule.isKnown() && OMPInModule) {
1486     OMPInModule.identifyKernels(M);
1487     return true;
1488   }
1489 
1490   return OMPInModule = false;
1491 }
1492 
1493 char OpenMPOptLegacyPass::ID = 0;
1494 
1495 INITIALIZE_PASS_BEGIN(OpenMPOptLegacyPass, "openmpopt",
1496                       "OpenMP specific optimizations", false, false)
1497 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
1498 INITIALIZE_PASS_END(OpenMPOptLegacyPass, "openmpopt",
1499                     "OpenMP specific optimizations", false, false)
1500 
1501 Pass *llvm::createOpenMPOptLegacyPass() { return new OpenMPOptLegacyPass(); }
1502