1 //===- FunctionImport.cpp - ThinLTO Summary-based Function Import ---------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements Function import based on summaries.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/Transforms/IPO/FunctionImport.h"
14 #include "llvm/ADT/ArrayRef.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/StringMap.h"
20 #include "llvm/ADT/StringRef.h"
21 #include "llvm/ADT/StringSet.h"
22 #include "llvm/Bitcode/BitcodeReader.h"
23 #include "llvm/IR/AutoUpgrade.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalObject.h"
28 #include "llvm/IR/GlobalValue.h"
29 #include "llvm/IR/GlobalVariable.h"
30 #include "llvm/IR/Metadata.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/IR/ModuleSummaryIndex.h"
33 #include "llvm/IRReader/IRReader.h"
34 #include "llvm/InitializePasses.h"
35 #include "llvm/Linker/IRMover.h"
36 #include "llvm/Object/ModuleSymbolTable.h"
37 #include "llvm/Object/SymbolicFile.h"
38 #include "llvm/Pass.h"
39 #include "llvm/Support/Casting.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/Errc.h"
43 #include "llvm/Support/Error.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/FileSystem.h"
46 #include "llvm/Support/SourceMgr.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Transforms/IPO/Internalize.h"
49 #include "llvm/Transforms/Utils/Cloning.h"
50 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
51 #include "llvm/Transforms/Utils/ValueMapper.h"
52 #include <cassert>
53 #include <memory>
54 #include <set>
55 #include <string>
56 #include <system_error>
57 #include <tuple>
58 #include <utility>
59 
60 using namespace llvm;
61 
62 #define DEBUG_TYPE "function-import"
63 
64 STATISTIC(NumImportedFunctionsThinLink,
65           "Number of functions thin link decided to import");
66 STATISTIC(NumImportedHotFunctionsThinLink,
67           "Number of hot functions thin link decided to import");
68 STATISTIC(NumImportedCriticalFunctionsThinLink,
69           "Number of critical functions thin link decided to import");
70 STATISTIC(NumImportedGlobalVarsThinLink,
71           "Number of global variables thin link decided to import");
72 STATISTIC(NumImportedFunctions, "Number of functions imported in backend");
73 STATISTIC(NumImportedGlobalVars,
74           "Number of global variables imported in backend");
75 STATISTIC(NumImportedModules, "Number of modules imported from");
76 STATISTIC(NumDeadSymbols, "Number of dead stripped symbols in index");
77 STATISTIC(NumLiveSymbols, "Number of live symbols in index");
78 
79 /// Limit on instruction count of imported functions.
80 static cl::opt<unsigned> ImportInstrLimit(
81     "import-instr-limit", cl::init(100), cl::Hidden, cl::value_desc("N"),
82     cl::desc("Only import functions with less than N instructions"));
83 
84 static cl::opt<int> ImportCutoff(
85     "import-cutoff", cl::init(-1), cl::Hidden, cl::value_desc("N"),
86     cl::desc("Only import first N functions if N>=0 (default -1)"));
87 
88 static cl::opt<bool>
89     ForceImportAll("force-import-all", cl::init(false), cl::Hidden,
90                    cl::desc("Import functions with noinline attribute"));
91 
92 static cl::opt<float>
93     ImportInstrFactor("import-instr-evolution-factor", cl::init(0.7),
94                       cl::Hidden, cl::value_desc("x"),
95                       cl::desc("As we import functions, multiply the "
96                                "`import-instr-limit` threshold by this factor "
97                                "before processing newly imported functions"));
98 
99 static cl::opt<float> ImportHotInstrFactor(
100     "import-hot-evolution-factor", cl::init(1.0), cl::Hidden,
101     cl::value_desc("x"),
102     cl::desc("As we import functions called from hot callsite, multiply the "
103              "`import-instr-limit` threshold by this factor "
104              "before processing newly imported functions"));
105 
106 static cl::opt<float> ImportHotMultiplier(
107     "import-hot-multiplier", cl::init(10.0), cl::Hidden, cl::value_desc("x"),
108     cl::desc("Multiply the `import-instr-limit` threshold for hot callsites"));
109 
110 static cl::opt<float> ImportCriticalMultiplier(
111     "import-critical-multiplier", cl::init(100.0), cl::Hidden,
112     cl::value_desc("x"),
113     cl::desc(
114         "Multiply the `import-instr-limit` threshold for critical callsites"));
115 
116 // FIXME: This multiplier was not really tuned up.
117 static cl::opt<float> ImportColdMultiplier(
118     "import-cold-multiplier", cl::init(0), cl::Hidden, cl::value_desc("N"),
119     cl::desc("Multiply the `import-instr-limit` threshold for cold callsites"));
120 
121 static cl::opt<bool> PrintImports("print-imports", cl::init(false), cl::Hidden,
122                                   cl::desc("Print imported functions"));
123 
124 static cl::opt<bool> PrintImportFailures(
125     "print-import-failures", cl::init(false), cl::Hidden,
126     cl::desc("Print information for functions rejected for importing"));
127 
128 static cl::opt<bool> ComputeDead("compute-dead", cl::init(true), cl::Hidden,
129                                  cl::desc("Compute dead symbols"));
130 
131 static cl::opt<bool> EnableImportMetadata(
132     "enable-import-metadata", cl::init(false), cl::Hidden,
133     cl::desc("Enable import metadata like 'thinlto_src_module'"));
134 
135 /// Summary file to use for function importing when using -function-import from
136 /// the command line.
137 static cl::opt<std::string>
138     SummaryFile("summary-file",
139                 cl::desc("The summary file to use for function importing."));
140 
141 /// Used when testing importing from distributed indexes via opt
142 // -function-import.
143 static cl::opt<bool>
144     ImportAllIndex("import-all-index",
145                    cl::desc("Import all external functions in index."));
146 
147 // Load lazily a module from \p FileName in \p Context.
148 static std::unique_ptr<Module> loadFile(const std::string &FileName,
149                                         LLVMContext &Context) {
150   SMDiagnostic Err;
151   LLVM_DEBUG(dbgs() << "Loading '" << FileName << "'\n");
152   // Metadata isn't loaded until functions are imported, to minimize
153   // the memory overhead.
154   std::unique_ptr<Module> Result =
155       getLazyIRFileModule(FileName, Err, Context,
156                           /* ShouldLazyLoadMetadata = */ true);
157   if (!Result) {
158     Err.print("function-import", errs());
159     report_fatal_error("Abort");
160   }
161 
162   return Result;
163 }
164 
165 /// Given a list of possible callee implementation for a call site, select one
166 /// that fits the \p Threshold.
167 ///
168 /// FIXME: select "best" instead of first that fits. But what is "best"?
169 /// - The smallest: more likely to be inlined.
170 /// - The one with the least outgoing edges (already well optimized).
171 /// - One from a module already being imported from in order to reduce the
172 ///   number of source modules parsed/linked.
173 /// - One that has PGO data attached.
174 /// - [insert you fancy metric here]
175 static const GlobalValueSummary *
176 selectCallee(const ModuleSummaryIndex &Index,
177              ArrayRef<std::unique_ptr<GlobalValueSummary>> CalleeSummaryList,
178              unsigned Threshold, StringRef CallerModulePath,
179              FunctionImporter::ImportFailureReason &Reason,
180              GlobalValue::GUID GUID) {
181   Reason = FunctionImporter::ImportFailureReason::None;
182   auto It = llvm::find_if(
183       CalleeSummaryList,
184       [&](const std::unique_ptr<GlobalValueSummary> &SummaryPtr) {
185         auto *GVSummary = SummaryPtr.get();
186         if (!Index.isGlobalValueLive(GVSummary)) {
187           Reason = FunctionImporter::ImportFailureReason::NotLive;
188           return false;
189         }
190 
191         if (GlobalValue::isInterposableLinkage(GVSummary->linkage())) {
192           Reason = FunctionImporter::ImportFailureReason::InterposableLinkage;
193           // There is no point in importing these, we can't inline them
194           return false;
195         }
196 
197         auto *Summary = cast<FunctionSummary>(GVSummary->getBaseObject());
198 
199         // If this is a local function, make sure we import the copy
200         // in the caller's module. The only time a local function can
201         // share an entry in the index is if there is a local with the same name
202         // in another module that had the same source file name (in a different
203         // directory), where each was compiled in their own directory so there
204         // was not distinguishing path.
205         // However, do the import from another module if there is only one
206         // entry in the list - in that case this must be a reference due
207         // to indirect call profile data, since a function pointer can point to
208         // a local in another module.
209         if (GlobalValue::isLocalLinkage(Summary->linkage()) &&
210             CalleeSummaryList.size() > 1 &&
211             Summary->modulePath() != CallerModulePath) {
212           Reason =
213               FunctionImporter::ImportFailureReason::LocalLinkageNotInModule;
214           return false;
215         }
216 
217         if ((Summary->instCount() > Threshold) &&
218             !Summary->fflags().AlwaysInline && !ForceImportAll) {
219           Reason = FunctionImporter::ImportFailureReason::TooLarge;
220           return false;
221         }
222 
223         // Skip if it isn't legal to import (e.g. may reference unpromotable
224         // locals).
225         if (Summary->notEligibleToImport()) {
226           Reason = FunctionImporter::ImportFailureReason::NotEligible;
227           return false;
228         }
229 
230         // Don't bother importing if we can't inline it anyway.
231         if (Summary->fflags().NoInline && !ForceImportAll) {
232           Reason = FunctionImporter::ImportFailureReason::NoInline;
233           return false;
234         }
235 
236         return true;
237       });
238   if (It == CalleeSummaryList.end())
239     return nullptr;
240 
241   return cast<GlobalValueSummary>(It->get());
242 }
243 
244 namespace {
245 
246 using EdgeInfo =
247     std::tuple<const GlobalValueSummary *, unsigned /* Threshold */>;
248 
249 } // anonymous namespace
250 
251 static bool shouldImportGlobal(const ValueInfo &VI,
252                                const GVSummaryMapTy &DefinedGVSummaries) {
253   const auto &GVS = DefinedGVSummaries.find(VI.getGUID());
254   if (GVS == DefinedGVSummaries.end())
255     return true;
256   // We should not skip import if the module contains a definition with
257   // interposable linkage type. This is required for correctness in
258   // the situation with two following conditions:
259   // * the def with interposable linkage is non-prevailing,
260   // * there is a prevailing def available for import and marked read-only.
261   // In this case, the non-prevailing def will be converted to a declaration,
262   // while the prevailing one becomes internal, thus no definitions will be
263   // available for linking. In order to prevent undefined symbol link error,
264   // the prevailing definition must be imported.
265   // FIXME: Consider adding a check that the suitable prevailing definition
266   // exists and marked read-only.
267   if (VI.getSummaryList().size() > 1 &&
268       GlobalValue::isInterposableLinkage(GVS->second->linkage()))
269     return true;
270 
271   return false;
272 }
273 
274 static void computeImportForReferencedGlobals(
275     const GlobalValueSummary &Summary, const ModuleSummaryIndex &Index,
276     const GVSummaryMapTy &DefinedGVSummaries,
277     SmallVectorImpl<EdgeInfo> &Worklist,
278     FunctionImporter::ImportMapTy &ImportList,
279     StringMap<FunctionImporter::ExportSetTy> *ExportLists) {
280   for (auto &VI : Summary.refs()) {
281     if (!shouldImportGlobal(VI, DefinedGVSummaries)) {
282       LLVM_DEBUG(
283           dbgs() << "Ref ignored! Target already in destination module.\n");
284       continue;
285     }
286 
287     LLVM_DEBUG(dbgs() << " ref -> " << VI << "\n");
288 
289     // If this is a local variable, make sure we import the copy
290     // in the caller's module. The only time a local variable can
291     // share an entry in the index is if there is a local with the same name
292     // in another module that had the same source file name (in a different
293     // directory), where each was compiled in their own directory so there
294     // was not distinguishing path.
295     auto LocalNotInModule = [&](const GlobalValueSummary *RefSummary) -> bool {
296       return GlobalValue::isLocalLinkage(RefSummary->linkage()) &&
297              RefSummary->modulePath() != Summary.modulePath();
298     };
299 
300     for (auto &RefSummary : VI.getSummaryList())
301       if (isa<GlobalVarSummary>(RefSummary.get()) &&
302           Index.canImportGlobalVar(RefSummary.get(), /* AnalyzeRefs */ true) &&
303           !LocalNotInModule(RefSummary.get())) {
304         auto ILI = ImportList[RefSummary->modulePath()].insert(VI.getGUID());
305         // Only update stat and exports if we haven't already imported this
306         // variable.
307         if (!ILI.second)
308           break;
309         NumImportedGlobalVarsThinLink++;
310         // Any references made by this variable will be marked exported later,
311         // in ComputeCrossModuleImport, after import decisions are complete,
312         // which is more efficient than adding them here.
313         if (ExportLists)
314           (*ExportLists)[RefSummary->modulePath()].insert(VI);
315 
316         // If variable is not writeonly we attempt to recursively analyze
317         // its references in order to import referenced constants.
318         if (!Index.isWriteOnly(cast<GlobalVarSummary>(RefSummary.get())))
319           Worklist.emplace_back(RefSummary.get(), 0);
320         break;
321       }
322   }
323 }
324 
325 static const char *
326 getFailureName(FunctionImporter::ImportFailureReason Reason) {
327   switch (Reason) {
328   case FunctionImporter::ImportFailureReason::None:
329     return "None";
330   case FunctionImporter::ImportFailureReason::GlobalVar:
331     return "GlobalVar";
332   case FunctionImporter::ImportFailureReason::NotLive:
333     return "NotLive";
334   case FunctionImporter::ImportFailureReason::TooLarge:
335     return "TooLarge";
336   case FunctionImporter::ImportFailureReason::InterposableLinkage:
337     return "InterposableLinkage";
338   case FunctionImporter::ImportFailureReason::LocalLinkageNotInModule:
339     return "LocalLinkageNotInModule";
340   case FunctionImporter::ImportFailureReason::NotEligible:
341     return "NotEligible";
342   case FunctionImporter::ImportFailureReason::NoInline:
343     return "NoInline";
344   }
345   llvm_unreachable("invalid reason");
346 }
347 
348 /// Compute the list of functions to import for a given caller. Mark these
349 /// imported functions and the symbols they reference in their source module as
350 /// exported from their source module.
351 static void computeImportForFunction(
352     const FunctionSummary &Summary, const ModuleSummaryIndex &Index,
353     const unsigned Threshold, const GVSummaryMapTy &DefinedGVSummaries,
354     SmallVectorImpl<EdgeInfo> &Worklist,
355     FunctionImporter::ImportMapTy &ImportList,
356     StringMap<FunctionImporter::ExportSetTy> *ExportLists,
357     FunctionImporter::ImportThresholdsTy &ImportThresholds) {
358   computeImportForReferencedGlobals(Summary, Index, DefinedGVSummaries,
359                                     Worklist, ImportList, ExportLists);
360   static int ImportCount = 0;
361   for (auto &Edge : Summary.calls()) {
362     ValueInfo VI = Edge.first;
363     LLVM_DEBUG(dbgs() << " edge -> " << VI << " Threshold:" << Threshold
364                       << "\n");
365 
366     if (ImportCutoff >= 0 && ImportCount >= ImportCutoff) {
367       LLVM_DEBUG(dbgs() << "ignored! import-cutoff value of " << ImportCutoff
368                         << " reached.\n");
369       continue;
370     }
371 
372     if (DefinedGVSummaries.count(VI.getGUID())) {
373       // FIXME: Consider not skipping import if the module contains
374       // a non-prevailing def with interposable linkage. The prevailing copy
375       // can safely be imported (see shouldImportGlobal()).
376       LLVM_DEBUG(dbgs() << "ignored! Target already in destination module.\n");
377       continue;
378     }
379 
380     auto GetBonusMultiplier = [](CalleeInfo::HotnessType Hotness) -> float {
381       if (Hotness == CalleeInfo::HotnessType::Hot)
382         return ImportHotMultiplier;
383       if (Hotness == CalleeInfo::HotnessType::Cold)
384         return ImportColdMultiplier;
385       if (Hotness == CalleeInfo::HotnessType::Critical)
386         return ImportCriticalMultiplier;
387       return 1.0;
388     };
389 
390     const auto NewThreshold =
391         Threshold * GetBonusMultiplier(Edge.second.getHotness());
392 
393     auto IT = ImportThresholds.insert(std::make_pair(
394         VI.getGUID(), std::make_tuple(NewThreshold, nullptr, nullptr)));
395     bool PreviouslyVisited = !IT.second;
396     auto &ProcessedThreshold = std::get<0>(IT.first->second);
397     auto &CalleeSummary = std::get<1>(IT.first->second);
398     auto &FailureInfo = std::get<2>(IT.first->second);
399 
400     bool IsHotCallsite =
401         Edge.second.getHotness() == CalleeInfo::HotnessType::Hot;
402     bool IsCriticalCallsite =
403         Edge.second.getHotness() == CalleeInfo::HotnessType::Critical;
404 
405     const FunctionSummary *ResolvedCalleeSummary = nullptr;
406     if (CalleeSummary) {
407       assert(PreviouslyVisited);
408       // Since the traversal of the call graph is DFS, we can revisit a function
409       // a second time with a higher threshold. In this case, it is added back
410       // to the worklist with the new threshold (so that its own callee chains
411       // can be considered with the higher threshold).
412       if (NewThreshold <= ProcessedThreshold) {
413         LLVM_DEBUG(
414             dbgs() << "ignored! Target was already imported with Threshold "
415                    << ProcessedThreshold << "\n");
416         continue;
417       }
418       // Update with new larger threshold.
419       ProcessedThreshold = NewThreshold;
420       ResolvedCalleeSummary = cast<FunctionSummary>(CalleeSummary);
421     } else {
422       // If we already rejected importing a callee at the same or higher
423       // threshold, don't waste time calling selectCallee.
424       if (PreviouslyVisited && NewThreshold <= ProcessedThreshold) {
425         LLVM_DEBUG(
426             dbgs() << "ignored! Target was already rejected with Threshold "
427             << ProcessedThreshold << "\n");
428         if (PrintImportFailures) {
429           assert(FailureInfo &&
430                  "Expected FailureInfo for previously rejected candidate");
431           FailureInfo->Attempts++;
432         }
433         continue;
434       }
435 
436       FunctionImporter::ImportFailureReason Reason;
437       CalleeSummary = selectCallee(Index, VI.getSummaryList(), NewThreshold,
438                                    Summary.modulePath(), Reason, VI.getGUID());
439       if (!CalleeSummary) {
440         // Update with new larger threshold if this was a retry (otherwise
441         // we would have already inserted with NewThreshold above). Also
442         // update failure info if requested.
443         if (PreviouslyVisited) {
444           ProcessedThreshold = NewThreshold;
445           if (PrintImportFailures) {
446             assert(FailureInfo &&
447                    "Expected FailureInfo for previously rejected candidate");
448             FailureInfo->Reason = Reason;
449             FailureInfo->Attempts++;
450             FailureInfo->MaxHotness =
451                 std::max(FailureInfo->MaxHotness, Edge.second.getHotness());
452           }
453         } else if (PrintImportFailures) {
454           assert(!FailureInfo &&
455                  "Expected no FailureInfo for newly rejected candidate");
456           FailureInfo = std::make_unique<FunctionImporter::ImportFailureInfo>(
457               VI, Edge.second.getHotness(), Reason, 1);
458         }
459         if (ForceImportAll) {
460           std::string Msg = std::string("Failed to import function ") +
461                             VI.name().str() + " due to " +
462                             getFailureName(Reason);
463           auto Error = make_error<StringError>(
464               Msg, make_error_code(errc::not_supported));
465           logAllUnhandledErrors(std::move(Error), errs(),
466                                 "Error importing module: ");
467           break;
468         } else {
469           LLVM_DEBUG(dbgs()
470                      << "ignored! No qualifying callee with summary found.\n");
471           continue;
472         }
473       }
474 
475       // "Resolve" the summary
476       CalleeSummary = CalleeSummary->getBaseObject();
477       ResolvedCalleeSummary = cast<FunctionSummary>(CalleeSummary);
478 
479       assert((ResolvedCalleeSummary->fflags().AlwaysInline || ForceImportAll ||
480               (ResolvedCalleeSummary->instCount() <= NewThreshold)) &&
481              "selectCallee() didn't honor the threshold");
482 
483       auto ExportModulePath = ResolvedCalleeSummary->modulePath();
484       auto ILI = ImportList[ExportModulePath].insert(VI.getGUID());
485       // We previously decided to import this GUID definition if it was already
486       // inserted in the set of imports from the exporting module.
487       bool PreviouslyImported = !ILI.second;
488       if (!PreviouslyImported) {
489         NumImportedFunctionsThinLink++;
490         if (IsHotCallsite)
491           NumImportedHotFunctionsThinLink++;
492         if (IsCriticalCallsite)
493           NumImportedCriticalFunctionsThinLink++;
494       }
495 
496       // Any calls/references made by this function will be marked exported
497       // later, in ComputeCrossModuleImport, after import decisions are
498       // complete, which is more efficient than adding them here.
499       if (ExportLists)
500         (*ExportLists)[ExportModulePath].insert(VI);
501     }
502 
503     auto GetAdjustedThreshold = [](unsigned Threshold, bool IsHotCallsite) {
504       // Adjust the threshold for next level of imported functions.
505       // The threshold is different for hot callsites because we can then
506       // inline chains of hot calls.
507       if (IsHotCallsite)
508         return Threshold * ImportHotInstrFactor;
509       return Threshold * ImportInstrFactor;
510     };
511 
512     const auto AdjThreshold = GetAdjustedThreshold(Threshold, IsHotCallsite);
513 
514     ImportCount++;
515 
516     // Insert the newly imported function to the worklist.
517     Worklist.emplace_back(ResolvedCalleeSummary, AdjThreshold);
518   }
519 }
520 
521 /// Given the list of globals defined in a module, compute the list of imports
522 /// as well as the list of "exports", i.e. the list of symbols referenced from
523 /// another module (that may require promotion).
524 static void ComputeImportForModule(
525     const GVSummaryMapTy &DefinedGVSummaries, const ModuleSummaryIndex &Index,
526     StringRef ModName, FunctionImporter::ImportMapTy &ImportList,
527     StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr) {
528   // Worklist contains the list of function imported in this module, for which
529   // we will analyse the callees and may import further down the callgraph.
530   SmallVector<EdgeInfo, 128> Worklist;
531   FunctionImporter::ImportThresholdsTy ImportThresholds;
532 
533   // Populate the worklist with the import for the functions in the current
534   // module
535   for (auto &GVSummary : DefinedGVSummaries) {
536 #ifndef NDEBUG
537     // FIXME: Change the GVSummaryMapTy to hold ValueInfo instead of GUID
538     // so this map look up (and possibly others) can be avoided.
539     auto VI = Index.getValueInfo(GVSummary.first);
540 #endif
541     if (!Index.isGlobalValueLive(GVSummary.second)) {
542       LLVM_DEBUG(dbgs() << "Ignores Dead GUID: " << VI << "\n");
543       continue;
544     }
545     auto *FuncSummary =
546         dyn_cast<FunctionSummary>(GVSummary.second->getBaseObject());
547     if (!FuncSummary)
548       // Skip import for global variables
549       continue;
550     LLVM_DEBUG(dbgs() << "Initialize import for " << VI << "\n");
551     computeImportForFunction(*FuncSummary, Index, ImportInstrLimit,
552                              DefinedGVSummaries, Worklist, ImportList,
553                              ExportLists, ImportThresholds);
554   }
555 
556   // Process the newly imported functions and add callees to the worklist.
557   while (!Worklist.empty()) {
558     auto GVInfo = Worklist.pop_back_val();
559     auto *Summary = std::get<0>(GVInfo);
560     auto Threshold = std::get<1>(GVInfo);
561 
562     if (auto *FS = dyn_cast<FunctionSummary>(Summary))
563       computeImportForFunction(*FS, Index, Threshold, DefinedGVSummaries,
564                                Worklist, ImportList, ExportLists,
565                                ImportThresholds);
566     else
567       computeImportForReferencedGlobals(*Summary, Index, DefinedGVSummaries,
568                                         Worklist, ImportList, ExportLists);
569   }
570 
571   // Print stats about functions considered but rejected for importing
572   // when requested.
573   if (PrintImportFailures) {
574     dbgs() << "Missed imports into module " << ModName << "\n";
575     for (auto &I : ImportThresholds) {
576       auto &ProcessedThreshold = std::get<0>(I.second);
577       auto &CalleeSummary = std::get<1>(I.second);
578       auto &FailureInfo = std::get<2>(I.second);
579       if (CalleeSummary)
580         continue; // We are going to import.
581       assert(FailureInfo);
582       FunctionSummary *FS = nullptr;
583       if (!FailureInfo->VI.getSummaryList().empty())
584         FS = dyn_cast<FunctionSummary>(
585             FailureInfo->VI.getSummaryList()[0]->getBaseObject());
586       dbgs() << FailureInfo->VI
587              << ": Reason = " << getFailureName(FailureInfo->Reason)
588              << ", Threshold = " << ProcessedThreshold
589              << ", Size = " << (FS ? (int)FS->instCount() : -1)
590              << ", MaxHotness = " << getHotnessName(FailureInfo->MaxHotness)
591              << ", Attempts = " << FailureInfo->Attempts << "\n";
592     }
593   }
594 }
595 
596 #ifndef NDEBUG
597 static bool isGlobalVarSummary(const ModuleSummaryIndex &Index, ValueInfo VI) {
598   auto SL = VI.getSummaryList();
599   return SL.empty()
600              ? false
601              : SL[0]->getSummaryKind() == GlobalValueSummary::GlobalVarKind;
602 }
603 
604 static bool isGlobalVarSummary(const ModuleSummaryIndex &Index,
605                                GlobalValue::GUID G) {
606   if (const auto &VI = Index.getValueInfo(G))
607     return isGlobalVarSummary(Index, VI);
608   return false;
609 }
610 
611 template <class T>
612 static unsigned numGlobalVarSummaries(const ModuleSummaryIndex &Index,
613                                       T &Cont) {
614   unsigned NumGVS = 0;
615   for (auto &V : Cont)
616     if (isGlobalVarSummary(Index, V))
617       ++NumGVS;
618   return NumGVS;
619 }
620 #endif
621 
622 #ifndef NDEBUG
623 static bool
624 checkVariableImport(const ModuleSummaryIndex &Index,
625                     StringMap<FunctionImporter::ImportMapTy> &ImportLists,
626                     StringMap<FunctionImporter::ExportSetTy> &ExportLists) {
627 
628   DenseSet<GlobalValue::GUID> FlattenedImports;
629 
630   for (auto &ImportPerModule : ImportLists)
631     for (auto &ExportPerModule : ImportPerModule.second)
632       FlattenedImports.insert(ExportPerModule.second.begin(),
633                               ExportPerModule.second.end());
634 
635   // Checks that all GUIDs of read/writeonly vars we see in export lists
636   // are also in the import lists. Otherwise we my face linker undefs,
637   // because readonly and writeonly vars are internalized in their
638   // source modules.
639   auto IsReadOrWriteOnlyVar = [&](StringRef ModulePath, const ValueInfo &VI) {
640     auto *GVS = dyn_cast_or_null<GlobalVarSummary>(
641         Index.findSummaryInModule(VI, ModulePath));
642     return GVS && (Index.isReadOnly(GVS) || Index.isWriteOnly(GVS));
643   };
644 
645   for (auto &ExportPerModule : ExportLists)
646     for (auto &VI : ExportPerModule.second)
647       if (!FlattenedImports.count(VI.getGUID()) &&
648           IsReadOrWriteOnlyVar(ExportPerModule.first(), VI))
649         return false;
650 
651   return true;
652 }
653 #endif
654 
655 /// Compute all the import and export for every module using the Index.
656 void llvm::ComputeCrossModuleImport(
657     const ModuleSummaryIndex &Index,
658     const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
659     StringMap<FunctionImporter::ImportMapTy> &ImportLists,
660     StringMap<FunctionImporter::ExportSetTy> &ExportLists) {
661   // For each module that has function defined, compute the import/export lists.
662   for (auto &DefinedGVSummaries : ModuleToDefinedGVSummaries) {
663     auto &ImportList = ImportLists[DefinedGVSummaries.first()];
664     LLVM_DEBUG(dbgs() << "Computing import for Module '"
665                       << DefinedGVSummaries.first() << "'\n");
666     ComputeImportForModule(DefinedGVSummaries.second, Index,
667                            DefinedGVSummaries.first(), ImportList,
668                            &ExportLists);
669   }
670 
671   // When computing imports we only added the variables and functions being
672   // imported to the export list. We also need to mark any references and calls
673   // they make as exported as well. We do this here, as it is more efficient
674   // since we may import the same values multiple times into different modules
675   // during the import computation.
676   for (auto &ELI : ExportLists) {
677     FunctionImporter::ExportSetTy NewExports;
678     const auto &DefinedGVSummaries =
679         ModuleToDefinedGVSummaries.lookup(ELI.first());
680     for (auto &EI : ELI.second) {
681       // Find the copy defined in the exporting module so that we can mark the
682       // values it references in that specific definition as exported.
683       // Below we will add all references and called values, without regard to
684       // whether they are also defined in this module. We subsequently prune the
685       // list to only include those defined in the exporting module, see comment
686       // there as to why.
687       auto DS = DefinedGVSummaries.find(EI.getGUID());
688       // Anything marked exported during the import computation must have been
689       // defined in the exporting module.
690       assert(DS != DefinedGVSummaries.end());
691       auto *S = DS->getSecond();
692       S = S->getBaseObject();
693       if (auto *GVS = dyn_cast<GlobalVarSummary>(S)) {
694         // Export referenced functions and variables. We don't export/promote
695         // objects referenced by writeonly variable initializer, because
696         // we convert such variables initializers to "zeroinitializer".
697         // See processGlobalForThinLTO.
698         if (!Index.isWriteOnly(GVS))
699           for (const auto &VI : GVS->refs())
700             NewExports.insert(VI);
701       } else {
702         auto *FS = cast<FunctionSummary>(S);
703         for (auto &Edge : FS->calls())
704           NewExports.insert(Edge.first);
705         for (auto &Ref : FS->refs())
706           NewExports.insert(Ref);
707       }
708     }
709     // Prune list computed above to only include values defined in the exporting
710     // module. We do this after the above insertion since we may hit the same
711     // ref/call target multiple times in above loop, and it is more efficient to
712     // avoid a set lookup each time.
713     for (auto EI = NewExports.begin(); EI != NewExports.end();) {
714       if (!DefinedGVSummaries.count(EI->getGUID()))
715         NewExports.erase(EI++);
716       else
717         ++EI;
718     }
719     ELI.second.insert(NewExports.begin(), NewExports.end());
720   }
721 
722   assert(checkVariableImport(Index, ImportLists, ExportLists));
723 #ifndef NDEBUG
724   LLVM_DEBUG(dbgs() << "Import/Export lists for " << ImportLists.size()
725                     << " modules:\n");
726   for (auto &ModuleImports : ImportLists) {
727     auto ModName = ModuleImports.first();
728     auto &Exports = ExportLists[ModName];
729     unsigned NumGVS = numGlobalVarSummaries(Index, Exports);
730     LLVM_DEBUG(dbgs() << "* Module " << ModName << " exports "
731                       << Exports.size() - NumGVS << " functions and " << NumGVS
732                       << " vars. Imports from " << ModuleImports.second.size()
733                       << " modules.\n");
734     for (auto &Src : ModuleImports.second) {
735       auto SrcModName = Src.first();
736       unsigned NumGVSPerMod = numGlobalVarSummaries(Index, Src.second);
737       LLVM_DEBUG(dbgs() << " - " << Src.second.size() - NumGVSPerMod
738                         << " functions imported from " << SrcModName << "\n");
739       LLVM_DEBUG(dbgs() << " - " << NumGVSPerMod
740                         << " global vars imported from " << SrcModName << "\n");
741     }
742   }
743 #endif
744 }
745 
746 #ifndef NDEBUG
747 static void dumpImportListForModule(const ModuleSummaryIndex &Index,
748                                     StringRef ModulePath,
749                                     FunctionImporter::ImportMapTy &ImportList) {
750   LLVM_DEBUG(dbgs() << "* Module " << ModulePath << " imports from "
751                     << ImportList.size() << " modules.\n");
752   for (auto &Src : ImportList) {
753     auto SrcModName = Src.first();
754     unsigned NumGVSPerMod = numGlobalVarSummaries(Index, Src.second);
755     LLVM_DEBUG(dbgs() << " - " << Src.second.size() - NumGVSPerMod
756                       << " functions imported from " << SrcModName << "\n");
757     LLVM_DEBUG(dbgs() << " - " << NumGVSPerMod << " vars imported from "
758                       << SrcModName << "\n");
759   }
760 }
761 #endif
762 
763 /// Compute all the imports for the given module in the Index.
764 void llvm::ComputeCrossModuleImportForModule(
765     StringRef ModulePath, const ModuleSummaryIndex &Index,
766     FunctionImporter::ImportMapTy &ImportList) {
767   // Collect the list of functions this module defines.
768   // GUID -> Summary
769   GVSummaryMapTy FunctionSummaryMap;
770   Index.collectDefinedFunctionsForModule(ModulePath, FunctionSummaryMap);
771 
772   // Compute the import list for this module.
773   LLVM_DEBUG(dbgs() << "Computing import for Module '" << ModulePath << "'\n");
774   ComputeImportForModule(FunctionSummaryMap, Index, ModulePath, ImportList);
775 
776 #ifndef NDEBUG
777   dumpImportListForModule(Index, ModulePath, ImportList);
778 #endif
779 }
780 
781 // Mark all external summaries in Index for import into the given module.
782 // Used for distributed builds using a distributed index.
783 void llvm::ComputeCrossModuleImportForModuleFromIndex(
784     StringRef ModulePath, const ModuleSummaryIndex &Index,
785     FunctionImporter::ImportMapTy &ImportList) {
786   for (auto &GlobalList : Index) {
787     // Ignore entries for undefined references.
788     if (GlobalList.second.SummaryList.empty())
789       continue;
790 
791     auto GUID = GlobalList.first;
792     assert(GlobalList.second.SummaryList.size() == 1 &&
793            "Expected individual combined index to have one summary per GUID");
794     auto &Summary = GlobalList.second.SummaryList[0];
795     // Skip the summaries for the importing module. These are included to
796     // e.g. record required linkage changes.
797     if (Summary->modulePath() == ModulePath)
798       continue;
799     // Add an entry to provoke importing by thinBackend.
800     ImportList[Summary->modulePath()].insert(GUID);
801   }
802 #ifndef NDEBUG
803   dumpImportListForModule(Index, ModulePath, ImportList);
804 #endif
805 }
806 
807 // For SamplePGO, the indirect call targets for local functions will
808 // have its original name annotated in profile. We try to find the
809 // corresponding PGOFuncName as the GUID, and fix up the edges
810 // accordingly.
811 void updateValueInfoForIndirectCalls(ModuleSummaryIndex &Index,
812                                      FunctionSummary *FS) {
813   for (auto &EI : FS->mutableCalls()) {
814     if (!EI.first.getSummaryList().empty())
815       continue;
816     auto GUID = Index.getGUIDFromOriginalID(EI.first.getGUID());
817     if (GUID == 0)
818       continue;
819     // Update the edge to point directly to the correct GUID.
820     auto VI = Index.getValueInfo(GUID);
821     if (llvm::any_of(
822             VI.getSummaryList(),
823             [&](const std::unique_ptr<GlobalValueSummary> &SummaryPtr) {
824               // The mapping from OriginalId to GUID may return a GUID
825               // that corresponds to a static variable. Filter it out here.
826               // This can happen when
827               // 1) There is a call to a library function which is not defined
828               // in the index.
829               // 2) There is a static variable with the  OriginalGUID identical
830               // to the GUID of the library function in 1);
831               // When this happens the static variable in 2) will be found,
832               // which needs to be filtered out.
833               return SummaryPtr->getSummaryKind() ==
834                      GlobalValueSummary::GlobalVarKind;
835             }))
836       continue;
837     EI.first = VI;
838   }
839 }
840 
841 void llvm::updateIndirectCalls(ModuleSummaryIndex &Index) {
842   for (const auto &Entry : Index) {
843     for (auto &S : Entry.second.SummaryList) {
844       if (auto *FS = dyn_cast<FunctionSummary>(S.get()))
845         updateValueInfoForIndirectCalls(Index, FS);
846     }
847   }
848 }
849 
850 void llvm::computeDeadSymbolsAndUpdateIndirectCalls(
851     ModuleSummaryIndex &Index,
852     const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
853     function_ref<PrevailingType(GlobalValue::GUID)> isPrevailing) {
854   assert(!Index.withGlobalValueDeadStripping());
855   if (!ComputeDead ||
856       // Don't do anything when nothing is live, this is friendly with tests.
857       GUIDPreservedSymbols.empty()) {
858     // Still need to update indirect calls.
859     updateIndirectCalls(Index);
860     return;
861   }
862   unsigned LiveSymbols = 0;
863   SmallVector<ValueInfo, 128> Worklist;
864   Worklist.reserve(GUIDPreservedSymbols.size() * 2);
865   for (auto GUID : GUIDPreservedSymbols) {
866     ValueInfo VI = Index.getValueInfo(GUID);
867     if (!VI)
868       continue;
869     for (auto &S : VI.getSummaryList())
870       S->setLive(true);
871   }
872 
873   // Add values flagged in the index as live roots to the worklist.
874   for (const auto &Entry : Index) {
875     auto VI = Index.getValueInfo(Entry);
876     for (auto &S : Entry.second.SummaryList) {
877       if (auto *FS = dyn_cast<FunctionSummary>(S.get()))
878         updateValueInfoForIndirectCalls(Index, FS);
879       if (S->isLive()) {
880         LLVM_DEBUG(dbgs() << "Live root: " << VI << "\n");
881         Worklist.push_back(VI);
882         ++LiveSymbols;
883         break;
884       }
885     }
886   }
887 
888   // Make value live and add it to the worklist if it was not live before.
889   auto visit = [&](ValueInfo VI, bool IsAliasee) {
890     // FIXME: If we knew which edges were created for indirect call profiles,
891     // we could skip them here. Any that are live should be reached via
892     // other edges, e.g. reference edges. Otherwise, using a profile collected
893     // on a slightly different binary might provoke preserving, importing
894     // and ultimately promoting calls to functions not linked into this
895     // binary, which increases the binary size unnecessarily. Note that
896     // if this code changes, the importer needs to change so that edges
897     // to functions marked dead are skipped.
898 
899     if (llvm::any_of(VI.getSummaryList(),
900                      [](const std::unique_ptr<llvm::GlobalValueSummary> &S) {
901                        return S->isLive();
902                      }))
903       return;
904 
905     // We only keep live symbols that are known to be non-prevailing if any are
906     // available_externally, linkonceodr, weakodr. Those symbols are discarded
907     // later in the EliminateAvailableExternally pass and setting them to
908     // not-live could break downstreams users of liveness information (PR36483)
909     // or limit optimization opportunities.
910     if (isPrevailing(VI.getGUID()) == PrevailingType::No) {
911       bool KeepAliveLinkage = false;
912       bool Interposable = false;
913       for (auto &S : VI.getSummaryList()) {
914         if (S->linkage() == GlobalValue::AvailableExternallyLinkage ||
915             S->linkage() == GlobalValue::WeakODRLinkage ||
916             S->linkage() == GlobalValue::LinkOnceODRLinkage)
917           KeepAliveLinkage = true;
918         else if (GlobalValue::isInterposableLinkage(S->linkage()))
919           Interposable = true;
920       }
921 
922       if (!IsAliasee) {
923         if (!KeepAliveLinkage)
924           return;
925 
926         if (Interposable)
927           report_fatal_error(
928               "Interposable and available_externally/linkonce_odr/weak_odr "
929               "symbol");
930       }
931     }
932 
933     for (auto &S : VI.getSummaryList())
934       S->setLive(true);
935     ++LiveSymbols;
936     Worklist.push_back(VI);
937   };
938 
939   while (!Worklist.empty()) {
940     auto VI = Worklist.pop_back_val();
941     for (auto &Summary : VI.getSummaryList()) {
942       if (auto *AS = dyn_cast<AliasSummary>(Summary.get())) {
943         // If this is an alias, visit the aliasee VI to ensure that all copies
944         // are marked live and it is added to the worklist for further
945         // processing of its references.
946         visit(AS->getAliaseeVI(), true);
947         continue;
948       }
949       for (auto Ref : Summary->refs())
950         visit(Ref, false);
951       if (auto *FS = dyn_cast<FunctionSummary>(Summary.get()))
952         for (auto Call : FS->calls())
953           visit(Call.first, false);
954     }
955   }
956   Index.setWithGlobalValueDeadStripping();
957 
958   unsigned DeadSymbols = Index.size() - LiveSymbols;
959   LLVM_DEBUG(dbgs() << LiveSymbols << " symbols Live, and " << DeadSymbols
960                     << " symbols Dead \n");
961   NumDeadSymbols += DeadSymbols;
962   NumLiveSymbols += LiveSymbols;
963 }
964 
965 // Compute dead symbols and propagate constants in combined index.
966 void llvm::computeDeadSymbolsWithConstProp(
967     ModuleSummaryIndex &Index,
968     const DenseSet<GlobalValue::GUID> &GUIDPreservedSymbols,
969     function_ref<PrevailingType(GlobalValue::GUID)> isPrevailing,
970     bool ImportEnabled) {
971   computeDeadSymbolsAndUpdateIndirectCalls(Index, GUIDPreservedSymbols,
972                                            isPrevailing);
973   if (ImportEnabled)
974     Index.propagateAttributes(GUIDPreservedSymbols);
975 }
976 
977 /// Compute the set of summaries needed for a ThinLTO backend compilation of
978 /// \p ModulePath.
979 void llvm::gatherImportedSummariesForModule(
980     StringRef ModulePath,
981     const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
982     const FunctionImporter::ImportMapTy &ImportList,
983     std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
984   // Include all summaries from the importing module.
985   ModuleToSummariesForIndex[std::string(ModulePath)] =
986       ModuleToDefinedGVSummaries.lookup(ModulePath);
987   // Include summaries for imports.
988   for (auto &ILI : ImportList) {
989     auto &SummariesForIndex =
990         ModuleToSummariesForIndex[std::string(ILI.first())];
991     const auto &DefinedGVSummaries =
992         ModuleToDefinedGVSummaries.lookup(ILI.first());
993     for (auto &GI : ILI.second) {
994       const auto &DS = DefinedGVSummaries.find(GI);
995       assert(DS != DefinedGVSummaries.end() &&
996              "Expected a defined summary for imported global value");
997       SummariesForIndex[GI] = DS->second;
998     }
999   }
1000 }
1001 
1002 /// Emit the files \p ModulePath will import from into \p OutputFilename.
1003 std::error_code llvm::EmitImportsFiles(
1004     StringRef ModulePath, StringRef OutputFilename,
1005     const std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
1006   std::error_code EC;
1007   raw_fd_ostream ImportsOS(OutputFilename, EC, sys::fs::OpenFlags::OF_None);
1008   if (EC)
1009     return EC;
1010   for (auto &ILI : ModuleToSummariesForIndex)
1011     // The ModuleToSummariesForIndex map includes an entry for the current
1012     // Module (needed for writing out the index files). We don't want to
1013     // include it in the imports file, however, so filter it out.
1014     if (ILI.first != ModulePath)
1015       ImportsOS << ILI.first << "\n";
1016   return std::error_code();
1017 }
1018 
1019 bool llvm::convertToDeclaration(GlobalValue &GV) {
1020   LLVM_DEBUG(dbgs() << "Converting to a declaration: `" << GV.getName()
1021                     << "\n");
1022   if (Function *F = dyn_cast<Function>(&GV)) {
1023     F->deleteBody();
1024     F->clearMetadata();
1025     F->setComdat(nullptr);
1026   } else if (GlobalVariable *V = dyn_cast<GlobalVariable>(&GV)) {
1027     V->setInitializer(nullptr);
1028     V->setLinkage(GlobalValue::ExternalLinkage);
1029     V->clearMetadata();
1030     V->setComdat(nullptr);
1031   } else {
1032     GlobalValue *NewGV;
1033     if (GV.getValueType()->isFunctionTy())
1034       NewGV =
1035           Function::Create(cast<FunctionType>(GV.getValueType()),
1036                            GlobalValue::ExternalLinkage, GV.getAddressSpace(),
1037                            "", GV.getParent());
1038     else
1039       NewGV =
1040           new GlobalVariable(*GV.getParent(), GV.getValueType(),
1041                              /*isConstant*/ false, GlobalValue::ExternalLinkage,
1042                              /*init*/ nullptr, "",
1043                              /*insertbefore*/ nullptr, GV.getThreadLocalMode(),
1044                              GV.getType()->getAddressSpace());
1045     NewGV->takeName(&GV);
1046     GV.replaceAllUsesWith(NewGV);
1047     return false;
1048   }
1049   if (!GV.isImplicitDSOLocal())
1050     GV.setDSOLocal(false);
1051   return true;
1052 }
1053 
1054 void llvm::thinLTOFinalizeInModule(Module &TheModule,
1055                                    const GVSummaryMapTy &DefinedGlobals,
1056                                    bool PropagateAttrs) {
1057   auto FinalizeInModule = [&](GlobalValue &GV, bool Propagate = false) {
1058     // See if the global summary analysis computed a new resolved linkage.
1059     const auto &GS = DefinedGlobals.find(GV.getGUID());
1060     if (GS == DefinedGlobals.end())
1061       return;
1062 
1063     if (Propagate)
1064       if (FunctionSummary *FS = dyn_cast<FunctionSummary>(GS->second)) {
1065         if (Function *F = dyn_cast<Function>(&GV)) {
1066           // TODO: propagate ReadNone and ReadOnly.
1067           if (FS->fflags().ReadNone && !F->doesNotAccessMemory())
1068             F->setDoesNotAccessMemory();
1069 
1070           if (FS->fflags().ReadOnly && !F->onlyReadsMemory())
1071             F->setOnlyReadsMemory();
1072 
1073           if (FS->fflags().NoRecurse && !F->doesNotRecurse())
1074             F->setDoesNotRecurse();
1075 
1076           if (FS->fflags().NoUnwind && !F->doesNotThrow())
1077             F->setDoesNotThrow();
1078         }
1079       }
1080 
1081     auto NewLinkage = GS->second->linkage();
1082     if (GlobalValue::isLocalLinkage(GV.getLinkage()) ||
1083         // Don't internalize anything here, because the code below
1084         // lacks necessary correctness checks. Leave this job to
1085         // LLVM 'internalize' pass.
1086         GlobalValue::isLocalLinkage(NewLinkage) ||
1087         // In case it was dead and already converted to declaration.
1088         GV.isDeclaration())
1089       return;
1090 
1091     // Set the potentially more constraining visibility computed from summaries.
1092     // The DefaultVisibility condition is because older GlobalValueSummary does
1093     // not record DefaultVisibility and we don't want to change protected/hidden
1094     // to default.
1095     if (GS->second->getVisibility() != GlobalValue::DefaultVisibility)
1096       GV.setVisibility(GS->second->getVisibility());
1097 
1098     if (NewLinkage == GV.getLinkage())
1099       return;
1100 
1101     // Check for a non-prevailing def that has interposable linkage
1102     // (e.g. non-odr weak or linkonce). In that case we can't simply
1103     // convert to available_externally, since it would lose the
1104     // interposable property and possibly get inlined. Simply drop
1105     // the definition in that case.
1106     if (GlobalValue::isAvailableExternallyLinkage(NewLinkage) &&
1107         GlobalValue::isInterposableLinkage(GV.getLinkage())) {
1108       if (!convertToDeclaration(GV))
1109         // FIXME: Change this to collect replaced GVs and later erase
1110         // them from the parent module once thinLTOResolvePrevailingGUID is
1111         // changed to enable this for aliases.
1112         llvm_unreachable("Expected GV to be converted");
1113     } else {
1114       // If all copies of the original symbol had global unnamed addr and
1115       // linkonce_odr linkage, it should be an auto hide symbol. In that case
1116       // the thin link would have marked it as CanAutoHide. Add hidden visibility
1117       // to the symbol to preserve the property.
1118       if (NewLinkage == GlobalValue::WeakODRLinkage &&
1119           GS->second->canAutoHide()) {
1120         assert(GV.hasLinkOnceODRLinkage() && GV.hasGlobalUnnamedAddr());
1121         GV.setVisibility(GlobalValue::HiddenVisibility);
1122       }
1123 
1124       LLVM_DEBUG(dbgs() << "ODR fixing up linkage for `" << GV.getName()
1125                         << "` from " << GV.getLinkage() << " to " << NewLinkage
1126                         << "\n");
1127       GV.setLinkage(NewLinkage);
1128     }
1129     // Remove declarations from comdats, including available_externally
1130     // as this is a declaration for the linker, and will be dropped eventually.
1131     // It is illegal for comdats to contain declarations.
1132     auto *GO = dyn_cast_or_null<GlobalObject>(&GV);
1133     if (GO && GO->isDeclarationForLinker() && GO->hasComdat())
1134       GO->setComdat(nullptr);
1135   };
1136 
1137   // Process functions and global now
1138   for (auto &GV : TheModule)
1139     FinalizeInModule(GV, PropagateAttrs);
1140   for (auto &GV : TheModule.globals())
1141     FinalizeInModule(GV);
1142   for (auto &GV : TheModule.aliases())
1143     FinalizeInModule(GV);
1144 }
1145 
1146 /// Run internalization on \p TheModule based on symmary analysis.
1147 void llvm::thinLTOInternalizeModule(Module &TheModule,
1148                                     const GVSummaryMapTy &DefinedGlobals) {
1149   // Declare a callback for the internalize pass that will ask for every
1150   // candidate GlobalValue if it can be internalized or not.
1151   auto MustPreserveGV = [&](const GlobalValue &GV) -> bool {
1152     // Lookup the linkage recorded in the summaries during global analysis.
1153     auto GS = DefinedGlobals.find(GV.getGUID());
1154     if (GS == DefinedGlobals.end()) {
1155       // Must have been promoted (possibly conservatively). Find original
1156       // name so that we can access the correct summary and see if it can
1157       // be internalized again.
1158       // FIXME: Eventually we should control promotion instead of promoting
1159       // and internalizing again.
1160       StringRef OrigName =
1161           ModuleSummaryIndex::getOriginalNameBeforePromote(GV.getName());
1162       std::string OrigId = GlobalValue::getGlobalIdentifier(
1163           OrigName, GlobalValue::InternalLinkage,
1164           TheModule.getSourceFileName());
1165       GS = DefinedGlobals.find(GlobalValue::getGUID(OrigId));
1166       if (GS == DefinedGlobals.end()) {
1167         // Also check the original non-promoted non-globalized name. In some
1168         // cases a preempted weak value is linked in as a local copy because
1169         // it is referenced by an alias (IRLinker::linkGlobalValueProto).
1170         // In that case, since it was originally not a local value, it was
1171         // recorded in the index using the original name.
1172         // FIXME: This may not be needed once PR27866 is fixed.
1173         GS = DefinedGlobals.find(GlobalValue::getGUID(OrigName));
1174         assert(GS != DefinedGlobals.end());
1175       }
1176     }
1177     return !GlobalValue::isLocalLinkage(GS->second->linkage());
1178   };
1179 
1180   // FIXME: See if we can just internalize directly here via linkage changes
1181   // based on the index, rather than invoking internalizeModule.
1182   internalizeModule(TheModule, MustPreserveGV);
1183 }
1184 
1185 /// Make alias a clone of its aliasee.
1186 static Function *replaceAliasWithAliasee(Module *SrcModule, GlobalAlias *GA) {
1187   Function *Fn = cast<Function>(GA->getAliaseeObject());
1188 
1189   ValueToValueMapTy VMap;
1190   Function *NewFn = CloneFunction(Fn, VMap);
1191   // Clone should use the original alias's linkage, visibility and name, and we
1192   // ensure all uses of alias instead use the new clone (casted if necessary).
1193   NewFn->setLinkage(GA->getLinkage());
1194   NewFn->setVisibility(GA->getVisibility());
1195   GA->replaceAllUsesWith(ConstantExpr::getBitCast(NewFn, GA->getType()));
1196   NewFn->takeName(GA);
1197   return NewFn;
1198 }
1199 
1200 // Internalize values that we marked with specific attribute
1201 // in processGlobalForThinLTO.
1202 static void internalizeGVsAfterImport(Module &M) {
1203   for (auto &GV : M.globals())
1204     // Skip GVs which have been converted to declarations
1205     // by dropDeadSymbols.
1206     if (!GV.isDeclaration() && GV.hasAttribute("thinlto-internalize")) {
1207       GV.setLinkage(GlobalValue::InternalLinkage);
1208       GV.setVisibility(GlobalValue::DefaultVisibility);
1209     }
1210 }
1211 
1212 // Automatically import functions in Module \p DestModule based on the summaries
1213 // index.
1214 Expected<bool> FunctionImporter::importFunctions(
1215     Module &DestModule, const FunctionImporter::ImportMapTy &ImportList) {
1216   LLVM_DEBUG(dbgs() << "Starting import for Module "
1217                     << DestModule.getModuleIdentifier() << "\n");
1218   unsigned ImportedCount = 0, ImportedGVCount = 0;
1219 
1220   IRMover Mover(DestModule);
1221   // Do the actual import of functions now, one Module at a time
1222   std::set<StringRef> ModuleNameOrderedList;
1223   for (auto &FunctionsToImportPerModule : ImportList) {
1224     ModuleNameOrderedList.insert(FunctionsToImportPerModule.first());
1225   }
1226   for (auto &Name : ModuleNameOrderedList) {
1227     // Get the module for the import
1228     const auto &FunctionsToImportPerModule = ImportList.find(Name);
1229     assert(FunctionsToImportPerModule != ImportList.end());
1230     Expected<std::unique_ptr<Module>> SrcModuleOrErr = ModuleLoader(Name);
1231     if (!SrcModuleOrErr)
1232       return SrcModuleOrErr.takeError();
1233     std::unique_ptr<Module> SrcModule = std::move(*SrcModuleOrErr);
1234     assert(&DestModule.getContext() == &SrcModule->getContext() &&
1235            "Context mismatch");
1236 
1237     // If modules were created with lazy metadata loading, materialize it
1238     // now, before linking it (otherwise this will be a noop).
1239     if (Error Err = SrcModule->materializeMetadata())
1240       return std::move(Err);
1241 
1242     auto &ImportGUIDs = FunctionsToImportPerModule->second;
1243     // Find the globals to import
1244     SetVector<GlobalValue *> GlobalsToImport;
1245     for (Function &F : *SrcModule) {
1246       if (!F.hasName())
1247         continue;
1248       auto GUID = F.getGUID();
1249       auto Import = ImportGUIDs.count(GUID);
1250       LLVM_DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing function "
1251                         << GUID << " " << F.getName() << " from "
1252                         << SrcModule->getSourceFileName() << "\n");
1253       if (Import) {
1254         if (Error Err = F.materialize())
1255           return std::move(Err);
1256         if (EnableImportMetadata) {
1257           // Add 'thinlto_src_module' metadata for statistics and debugging.
1258           F.setMetadata(
1259               "thinlto_src_module",
1260               MDNode::get(DestModule.getContext(),
1261                           {MDString::get(DestModule.getContext(),
1262                                          SrcModule->getSourceFileName())}));
1263         }
1264         GlobalsToImport.insert(&F);
1265       }
1266     }
1267     for (GlobalVariable &GV : SrcModule->globals()) {
1268       if (!GV.hasName())
1269         continue;
1270       auto GUID = GV.getGUID();
1271       auto Import = ImportGUIDs.count(GUID);
1272       LLVM_DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing global "
1273                         << GUID << " " << GV.getName() << " from "
1274                         << SrcModule->getSourceFileName() << "\n");
1275       if (Import) {
1276         if (Error Err = GV.materialize())
1277           return std::move(Err);
1278         ImportedGVCount += GlobalsToImport.insert(&GV);
1279       }
1280     }
1281     for (GlobalAlias &GA : SrcModule->aliases()) {
1282       if (!GA.hasName())
1283         continue;
1284       auto GUID = GA.getGUID();
1285       auto Import = ImportGUIDs.count(GUID);
1286       LLVM_DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing alias "
1287                         << GUID << " " << GA.getName() << " from "
1288                         << SrcModule->getSourceFileName() << "\n");
1289       if (Import) {
1290         if (Error Err = GA.materialize())
1291           return std::move(Err);
1292         // Import alias as a copy of its aliasee.
1293         GlobalObject *GO = GA.getAliaseeObject();
1294         if (Error Err = GO->materialize())
1295           return std::move(Err);
1296         auto *Fn = replaceAliasWithAliasee(SrcModule.get(), &GA);
1297         LLVM_DEBUG(dbgs() << "Is importing aliasee fn " << GO->getGUID() << " "
1298                           << GO->getName() << " from "
1299                           << SrcModule->getSourceFileName() << "\n");
1300         if (EnableImportMetadata) {
1301           // Add 'thinlto_src_module' metadata for statistics and debugging.
1302           Fn->setMetadata(
1303               "thinlto_src_module",
1304               MDNode::get(DestModule.getContext(),
1305                           {MDString::get(DestModule.getContext(),
1306                                          SrcModule->getSourceFileName())}));
1307         }
1308         GlobalsToImport.insert(Fn);
1309       }
1310     }
1311 
1312     // Upgrade debug info after we're done materializing all the globals and we
1313     // have loaded all the required metadata!
1314     UpgradeDebugInfo(*SrcModule);
1315 
1316     // Set the partial sample profile ratio in the profile summary module flag
1317     // of the imported source module, if applicable, so that the profile summary
1318     // module flag will match with that of the destination module when it's
1319     // imported.
1320     SrcModule->setPartialSampleProfileRatio(Index);
1321 
1322     // Link in the specified functions.
1323     if (renameModuleForThinLTO(*SrcModule, Index, ClearDSOLocalOnDeclarations,
1324                                &GlobalsToImport))
1325       return true;
1326 
1327     if (PrintImports) {
1328       for (const auto *GV : GlobalsToImport)
1329         dbgs() << DestModule.getSourceFileName() << ": Import " << GV->getName()
1330                << " from " << SrcModule->getSourceFileName() << "\n";
1331     }
1332 
1333     if (Error Err = Mover.move(
1334             std::move(SrcModule), GlobalsToImport.getArrayRef(),
1335             [](GlobalValue &, IRMover::ValueAdder) {},
1336             /*IsPerformingImport=*/true))
1337       report_fatal_error(Twine("Function Import: link error: ") +
1338                          toString(std::move(Err)));
1339 
1340     ImportedCount += GlobalsToImport.size();
1341     NumImportedModules++;
1342   }
1343 
1344   internalizeGVsAfterImport(DestModule);
1345 
1346   NumImportedFunctions += (ImportedCount - ImportedGVCount);
1347   NumImportedGlobalVars += ImportedGVCount;
1348 
1349   LLVM_DEBUG(dbgs() << "Imported " << ImportedCount - ImportedGVCount
1350                     << " functions for Module "
1351                     << DestModule.getModuleIdentifier() << "\n");
1352   LLVM_DEBUG(dbgs() << "Imported " << ImportedGVCount
1353                     << " global variables for Module "
1354                     << DestModule.getModuleIdentifier() << "\n");
1355   return ImportedCount;
1356 }
1357 
1358 static bool doImportingForModule(Module &M) {
1359   if (SummaryFile.empty())
1360     report_fatal_error("error: -function-import requires -summary-file\n");
1361   Expected<std::unique_ptr<ModuleSummaryIndex>> IndexPtrOrErr =
1362       getModuleSummaryIndexForFile(SummaryFile);
1363   if (!IndexPtrOrErr) {
1364     logAllUnhandledErrors(IndexPtrOrErr.takeError(), errs(),
1365                           "Error loading file '" + SummaryFile + "': ");
1366     return false;
1367   }
1368   std::unique_ptr<ModuleSummaryIndex> Index = std::move(*IndexPtrOrErr);
1369 
1370   // First step is collecting the import list.
1371   FunctionImporter::ImportMapTy ImportList;
1372   // If requested, simply import all functions in the index. This is used
1373   // when testing distributed backend handling via the opt tool, when
1374   // we have distributed indexes containing exactly the summaries to import.
1375   if (ImportAllIndex)
1376     ComputeCrossModuleImportForModuleFromIndex(M.getModuleIdentifier(), *Index,
1377                                                ImportList);
1378   else
1379     ComputeCrossModuleImportForModule(M.getModuleIdentifier(), *Index,
1380                                       ImportList);
1381 
1382   // Conservatively mark all internal values as promoted. This interface is
1383   // only used when doing importing via the function importing pass. The pass
1384   // is only enabled when testing importing via the 'opt' tool, which does
1385   // not do the ThinLink that would normally determine what values to promote.
1386   for (auto &I : *Index) {
1387     for (auto &S : I.second.SummaryList) {
1388       if (GlobalValue::isLocalLinkage(S->linkage()))
1389         S->setLinkage(GlobalValue::ExternalLinkage);
1390     }
1391   }
1392 
1393   // Next we need to promote to global scope and rename any local values that
1394   // are potentially exported to other modules.
1395   if (renameModuleForThinLTO(M, *Index, /*ClearDSOLocalOnDeclarations=*/false,
1396                              /*GlobalsToImport=*/nullptr)) {
1397     errs() << "Error renaming module\n";
1398     return false;
1399   }
1400 
1401   // Perform the import now.
1402   auto ModuleLoader = [&M](StringRef Identifier) {
1403     return loadFile(std::string(Identifier), M.getContext());
1404   };
1405   FunctionImporter Importer(*Index, ModuleLoader,
1406                             /*ClearDSOLocalOnDeclarations=*/false);
1407   Expected<bool> Result = Importer.importFunctions(M, ImportList);
1408 
1409   // FIXME: Probably need to propagate Errors through the pass manager.
1410   if (!Result) {
1411     logAllUnhandledErrors(Result.takeError(), errs(),
1412                           "Error importing module: ");
1413     return false;
1414   }
1415 
1416   return *Result;
1417 }
1418 
1419 namespace {
1420 
1421 /// Pass that performs cross-module function import provided a summary file.
1422 class FunctionImportLegacyPass : public ModulePass {
1423 public:
1424   /// Pass identification, replacement for typeid
1425   static char ID;
1426 
1427   explicit FunctionImportLegacyPass() : ModulePass(ID) {}
1428 
1429   /// Specify pass name for debug output
1430   StringRef getPassName() const override { return "Function Importing"; }
1431 
1432   bool runOnModule(Module &M) override {
1433     if (skipModule(M))
1434       return false;
1435 
1436     return doImportingForModule(M);
1437   }
1438 };
1439 
1440 } // end anonymous namespace
1441 
1442 PreservedAnalyses FunctionImportPass::run(Module &M,
1443                                           ModuleAnalysisManager &AM) {
1444   if (!doImportingForModule(M))
1445     return PreservedAnalyses::all();
1446 
1447   return PreservedAnalyses::none();
1448 }
1449 
1450 char FunctionImportLegacyPass::ID = 0;
1451 INITIALIZE_PASS(FunctionImportLegacyPass, "function-import",
1452                 "Summary Based Function Import", false, false)
1453 
1454 namespace llvm {
1455 
1456 Pass *createFunctionImportPass() {
1457   return new FunctionImportLegacyPass();
1458 }
1459 
1460 } // end namespace llvm
1461