1 //===- FunctionImport.cpp - ThinLTO Summary-based Function Import ---------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements Function import based on summaries.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Transforms/IPO/FunctionImport.h"
15
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/ADT/StringSet.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/IR/AutoUpgrade.h"
21 #include "llvm/IR/DiagnosticPrinter.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/IR/Module.h"
24 #include "llvm/IRReader/IRReader.h"
25 #include "llvm/Linker/Linker.h"
26 #include "llvm/Object/IRObjectFile.h"
27 #include "llvm/Object/ModuleSummaryIndexObjectFile.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/SourceMgr.h"
31 #include "llvm/Transforms/IPO/Internalize.h"
32 #include "llvm/Transforms/Utils/FunctionImportUtils.h"
33
34 #define DEBUG_TYPE "function-import"
35
36 using namespace llvm;
37
38 STATISTIC(NumImported, "Number of functions imported");
39
40 /// Limit on instruction count of imported functions.
41 static cl::opt<unsigned> ImportInstrLimit(
42 "import-instr-limit", cl::init(100), cl::Hidden, cl::value_desc("N"),
43 cl::desc("Only import functions with less than N instructions"));
44
45 static cl::opt<float>
46 ImportInstrFactor("import-instr-evolution-factor", cl::init(0.7),
47 cl::Hidden, cl::value_desc("x"),
48 cl::desc("As we import functions, multiply the "
49 "`import-instr-limit` threshold by this factor "
50 "before processing newly imported functions"));
51
52 static cl::opt<bool> PrintImports("print-imports", cl::init(false), cl::Hidden,
53 cl::desc("Print imported functions"));
54
55 // Temporary allows the function import pass to disable always linking
56 // referenced discardable symbols.
57 static cl::opt<bool>
58 DontForceImportReferencedDiscardableSymbols("disable-force-link-odr",
59 cl::init(false), cl::Hidden);
60
61 static cl::opt<bool> EnableImportMetadata(
62 "enable-import-metadata", cl::init(
63 #if !defined(NDEBUG)
64 true /*Enabled with asserts.*/
65 #else
66 false
67 #endif
68 ),
69 cl::Hidden, cl::desc("Enable import metadata like 'thinlto_src_module'"));
70
71 // Load lazily a module from \p FileName in \p Context.
loadFile(const std::string & FileName,LLVMContext & Context)72 static std::unique_ptr<Module> loadFile(const std::string &FileName,
73 LLVMContext &Context) {
74 SMDiagnostic Err;
75 DEBUG(dbgs() << "Loading '" << FileName << "'\n");
76 // Metadata isn't loaded until functions are imported, to minimize
77 // the memory overhead.
78 std::unique_ptr<Module> Result =
79 getLazyIRFileModule(FileName, Err, Context,
80 /* ShouldLazyLoadMetadata = */ true);
81 if (!Result) {
82 Err.print("function-import", errs());
83 report_fatal_error("Abort");
84 }
85
86 return Result;
87 }
88
89 namespace {
90
91 // Return true if the Summary describes a GlobalValue that can be externally
92 // referenced, i.e. it does not need renaming (linkage is not local) or renaming
93 // is possible (does not have a section for instance).
canBeExternallyReferenced(const GlobalValueSummary & Summary)94 static bool canBeExternallyReferenced(const GlobalValueSummary &Summary) {
95 if (!Summary.needsRenaming())
96 return true;
97
98 if (Summary.hasSection())
99 // Can't rename a global that needs renaming if has a section.
100 return false;
101
102 return true;
103 }
104
105 // Return true if \p GUID describes a GlobalValue that can be externally
106 // referenced, i.e. it does not need renaming (linkage is not local) or
107 // renaming is possible (does not have a section for instance).
canBeExternallyReferenced(const ModuleSummaryIndex & Index,GlobalValue::GUID GUID)108 static bool canBeExternallyReferenced(const ModuleSummaryIndex &Index,
109 GlobalValue::GUID GUID) {
110 auto Summaries = Index.findGlobalValueSummaryList(GUID);
111 if (Summaries == Index.end())
112 return true;
113 if (Summaries->second.size() != 1)
114 // If there are multiple globals with this GUID, then we know it is
115 // not a local symbol, and it is necessarily externally referenced.
116 return true;
117
118 // We don't need to check for the module path, because if it can't be
119 // externally referenced and we call it, it is necessarilly in the same
120 // module
121 return canBeExternallyReferenced(**Summaries->second.begin());
122 }
123
124 // Return true if the global described by \p Summary can be imported in another
125 // module.
eligibleForImport(const ModuleSummaryIndex & Index,const GlobalValueSummary & Summary)126 static bool eligibleForImport(const ModuleSummaryIndex &Index,
127 const GlobalValueSummary &Summary) {
128 if (!canBeExternallyReferenced(Summary))
129 // Can't import a global that needs renaming if has a section for instance.
130 // FIXME: we may be able to import it by copying it without promotion.
131 return false;
132
133 // Check references (and potential calls) in the same module. If the current
134 // value references a global that can't be externally referenced it is not
135 // eligible for import.
136 bool AllRefsCanBeExternallyReferenced =
137 llvm::all_of(Summary.refs(), [&](const ValueInfo &VI) {
138 return canBeExternallyReferenced(Index, VI.getGUID());
139 });
140 if (!AllRefsCanBeExternallyReferenced)
141 return false;
142
143 if (auto *FuncSummary = dyn_cast<FunctionSummary>(&Summary)) {
144 bool AllCallsCanBeExternallyReferenced = llvm::all_of(
145 FuncSummary->calls(), [&](const FunctionSummary::EdgeTy &Edge) {
146 return canBeExternallyReferenced(Index, Edge.first.getGUID());
147 });
148 if (!AllCallsCanBeExternallyReferenced)
149 return false;
150 }
151 return true;
152 }
153
154 /// Given a list of possible callee implementation for a call site, select one
155 /// that fits the \p Threshold.
156 ///
157 /// FIXME: select "best" instead of first that fits. But what is "best"?
158 /// - The smallest: more likely to be inlined.
159 /// - The one with the least outgoing edges (already well optimized).
160 /// - One from a module already being imported from in order to reduce the
161 /// number of source modules parsed/linked.
162 /// - One that has PGO data attached.
163 /// - [insert you fancy metric here]
164 static const GlobalValueSummary *
selectCallee(const ModuleSummaryIndex & Index,const GlobalValueSummaryList & CalleeSummaryList,unsigned Threshold)165 selectCallee(const ModuleSummaryIndex &Index,
166 const GlobalValueSummaryList &CalleeSummaryList,
167 unsigned Threshold) {
168 auto It = llvm::find_if(
169 CalleeSummaryList,
170 [&](const std::unique_ptr<GlobalValueSummary> &SummaryPtr) {
171 auto *GVSummary = SummaryPtr.get();
172 if (GlobalValue::isInterposableLinkage(GVSummary->linkage()))
173 // There is no point in importing these, we can't inline them
174 return false;
175 if (auto *AS = dyn_cast<AliasSummary>(GVSummary)) {
176 GVSummary = &AS->getAliasee();
177 // Alias can't point to "available_externally". However when we import
178 // linkOnceODR the linkage does not change. So we import the alias
179 // and aliasee only in this case.
180 // FIXME: we should import alias as available_externally *function*,
181 // the destination module does need to know it is an alias.
182 if (!GlobalValue::isLinkOnceODRLinkage(GVSummary->linkage()))
183 return false;
184 }
185
186 auto *Summary = cast<FunctionSummary>(GVSummary);
187
188 if (Summary->instCount() > Threshold)
189 return false;
190
191 if (!eligibleForImport(Index, *Summary))
192 return false;
193
194 return true;
195 });
196 if (It == CalleeSummaryList.end())
197 return nullptr;
198
199 return cast<GlobalValueSummary>(It->get());
200 }
201
202 /// Return the summary for the function \p GUID that fits the \p Threshold, or
203 /// null if there's no match.
selectCallee(GlobalValue::GUID GUID,unsigned Threshold,const ModuleSummaryIndex & Index)204 static const GlobalValueSummary *selectCallee(GlobalValue::GUID GUID,
205 unsigned Threshold,
206 const ModuleSummaryIndex &Index) {
207 auto CalleeSummaryList = Index.findGlobalValueSummaryList(GUID);
208 if (CalleeSummaryList == Index.end())
209 return nullptr; // This function does not have a summary
210 return selectCallee(Index, CalleeSummaryList->second, Threshold);
211 }
212
213 /// Mark the global \p GUID as export by module \p ExportModulePath if found in
214 /// this module. If it is a GlobalVariable, we also mark any referenced global
215 /// in the current module as exported.
exportGlobalInModule(const ModuleSummaryIndex & Index,StringRef ExportModulePath,GlobalValue::GUID GUID,FunctionImporter::ExportSetTy & ExportList)216 static void exportGlobalInModule(const ModuleSummaryIndex &Index,
217 StringRef ExportModulePath,
218 GlobalValue::GUID GUID,
219 FunctionImporter::ExportSetTy &ExportList) {
220 auto FindGlobalSummaryInModule =
221 [&](GlobalValue::GUID GUID) -> GlobalValueSummary *{
222 auto SummaryList = Index.findGlobalValueSummaryList(GUID);
223 if (SummaryList == Index.end())
224 // This global does not have a summary, it is not part of the ThinLTO
225 // process
226 return nullptr;
227 auto SummaryIter = llvm::find_if(
228 SummaryList->second,
229 [&](const std::unique_ptr<GlobalValueSummary> &Summary) {
230 return Summary->modulePath() == ExportModulePath;
231 });
232 if (SummaryIter == SummaryList->second.end())
233 return nullptr;
234 return SummaryIter->get();
235 };
236
237 auto *Summary = FindGlobalSummaryInModule(GUID);
238 if (!Summary)
239 return;
240 // We found it in the current module, mark as exported
241 ExportList.insert(GUID);
242
243 auto GVS = dyn_cast<GlobalVarSummary>(Summary);
244 if (!GVS)
245 return;
246 // FunctionImportGlobalProcessing::doPromoteLocalToGlobal() will always
247 // trigger importing the initializer for `constant unnamed addr` globals that
248 // are referenced. We conservatively export all the referenced symbols for
249 // every global to workaround this, so that the ExportList is accurate.
250 // FIXME: with a "isConstant" flag in the summary we could be more targetted.
251 for (auto &Ref : GVS->refs()) {
252 auto GUID = Ref.getGUID();
253 auto *RefSummary = FindGlobalSummaryInModule(GUID);
254 if (RefSummary)
255 // Found a ref in the current module, mark it as exported
256 ExportList.insert(GUID);
257 }
258 }
259
260 using EdgeInfo = std::pair<const FunctionSummary *, unsigned /* Threshold */>;
261
262 /// Compute the list of functions to import for a given caller. Mark these
263 /// imported functions and the symbols they reference in their source module as
264 /// exported from their source module.
computeImportForFunction(const FunctionSummary & Summary,const ModuleSummaryIndex & Index,unsigned Threshold,const GVSummaryMapTy & DefinedGVSummaries,SmallVectorImpl<EdgeInfo> & Worklist,FunctionImporter::ImportMapTy & ImportsForModule,StringMap<FunctionImporter::ExportSetTy> * ExportLists=nullptr)265 static void computeImportForFunction(
266 const FunctionSummary &Summary, const ModuleSummaryIndex &Index,
267 unsigned Threshold, const GVSummaryMapTy &DefinedGVSummaries,
268 SmallVectorImpl<EdgeInfo> &Worklist,
269 FunctionImporter::ImportMapTy &ImportsForModule,
270 StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr) {
271 for (auto &Edge : Summary.calls()) {
272 auto GUID = Edge.first.getGUID();
273 DEBUG(dbgs() << " edge -> " << GUID << " Threshold:" << Threshold << "\n");
274
275 if (DefinedGVSummaries.count(GUID)) {
276 DEBUG(dbgs() << "ignored! Target already in destination module.\n");
277 continue;
278 }
279
280 auto *CalleeSummary = selectCallee(GUID, Threshold, Index);
281 if (!CalleeSummary) {
282 DEBUG(dbgs() << "ignored! No qualifying callee with summary found.\n");
283 continue;
284 }
285 // "Resolve" the summary, traversing alias,
286 const FunctionSummary *ResolvedCalleeSummary;
287 if (isa<AliasSummary>(CalleeSummary)) {
288 ResolvedCalleeSummary = cast<FunctionSummary>(
289 &cast<AliasSummary>(CalleeSummary)->getAliasee());
290 assert(
291 GlobalValue::isLinkOnceODRLinkage(ResolvedCalleeSummary->linkage()) &&
292 "Unexpected alias to a non-linkonceODR in import list");
293 } else
294 ResolvedCalleeSummary = cast<FunctionSummary>(CalleeSummary);
295
296 assert(ResolvedCalleeSummary->instCount() <= Threshold &&
297 "selectCallee() didn't honor the threshold");
298
299 auto ExportModulePath = ResolvedCalleeSummary->modulePath();
300 auto &ProcessedThreshold = ImportsForModule[ExportModulePath][GUID];
301 /// Since the traversal of the call graph is DFS, we can revisit a function
302 /// a second time with a higher threshold. In this case, it is added back to
303 /// the worklist with the new threshold.
304 if (ProcessedThreshold && ProcessedThreshold >= Threshold) {
305 DEBUG(dbgs() << "ignored! Target was already seen with Threshold "
306 << ProcessedThreshold << "\n");
307 continue;
308 }
309 // Mark this function as imported in this module, with the current Threshold
310 ProcessedThreshold = Threshold;
311
312 // Make exports in the source module.
313 if (ExportLists) {
314 auto &ExportList = (*ExportLists)[ExportModulePath];
315 ExportList.insert(GUID);
316 // Mark all functions and globals referenced by this function as exported
317 // to the outside if they are defined in the same source module.
318 for (auto &Edge : ResolvedCalleeSummary->calls()) {
319 auto CalleeGUID = Edge.first.getGUID();
320 exportGlobalInModule(Index, ExportModulePath, CalleeGUID, ExportList);
321 }
322 for (auto &Ref : ResolvedCalleeSummary->refs()) {
323 auto GUID = Ref.getGUID();
324 exportGlobalInModule(Index, ExportModulePath, GUID, ExportList);
325 }
326 }
327
328 // Insert the newly imported function to the worklist.
329 Worklist.push_back(std::make_pair(ResolvedCalleeSummary, Threshold));
330 }
331 }
332
333 /// Given the list of globals defined in a module, compute the list of imports
334 /// as well as the list of "exports", i.e. the list of symbols referenced from
335 /// another module (that may require promotion).
ComputeImportForModule(const GVSummaryMapTy & DefinedGVSummaries,const ModuleSummaryIndex & Index,FunctionImporter::ImportMapTy & ImportsForModule,StringMap<FunctionImporter::ExportSetTy> * ExportLists=nullptr)336 static void ComputeImportForModule(
337 const GVSummaryMapTy &DefinedGVSummaries, const ModuleSummaryIndex &Index,
338 FunctionImporter::ImportMapTy &ImportsForModule,
339 StringMap<FunctionImporter::ExportSetTy> *ExportLists = nullptr) {
340 // Worklist contains the list of function imported in this module, for which
341 // we will analyse the callees and may import further down the callgraph.
342 SmallVector<EdgeInfo, 128> Worklist;
343
344 // Populate the worklist with the import for the functions in the current
345 // module
346 for (auto &GVSummary : DefinedGVSummaries) {
347 auto *Summary = GVSummary.second;
348 if (auto *AS = dyn_cast<AliasSummary>(Summary))
349 Summary = &AS->getAliasee();
350 auto *FuncSummary = dyn_cast<FunctionSummary>(Summary);
351 if (!FuncSummary)
352 // Skip import for global variables
353 continue;
354 DEBUG(dbgs() << "Initalize import for " << GVSummary.first << "\n");
355 computeImportForFunction(*FuncSummary, Index, ImportInstrLimit,
356 DefinedGVSummaries, Worklist, ImportsForModule,
357 ExportLists);
358 }
359
360 while (!Worklist.empty()) {
361 auto FuncInfo = Worklist.pop_back_val();
362 auto *Summary = FuncInfo.first;
363 auto Threshold = FuncInfo.second;
364
365 // Process the newly imported functions and add callees to the worklist.
366 // Adjust the threshold
367 Threshold = Threshold * ImportInstrFactor;
368
369 computeImportForFunction(*Summary, Index, Threshold, DefinedGVSummaries,
370 Worklist, ImportsForModule, ExportLists);
371 }
372 }
373
374 } // anonymous namespace
375
376 /// Compute all the import and export for every module using the Index.
ComputeCrossModuleImport(const ModuleSummaryIndex & Index,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries,StringMap<FunctionImporter::ImportMapTy> & ImportLists,StringMap<FunctionImporter::ExportSetTy> & ExportLists)377 void llvm::ComputeCrossModuleImport(
378 const ModuleSummaryIndex &Index,
379 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
380 StringMap<FunctionImporter::ImportMapTy> &ImportLists,
381 StringMap<FunctionImporter::ExportSetTy> &ExportLists) {
382 // For each module that has function defined, compute the import/export lists.
383 for (auto &DefinedGVSummaries : ModuleToDefinedGVSummaries) {
384 auto &ImportsForModule = ImportLists[DefinedGVSummaries.first()];
385 DEBUG(dbgs() << "Computing import for Module '"
386 << DefinedGVSummaries.first() << "'\n");
387 ComputeImportForModule(DefinedGVSummaries.second, Index, ImportsForModule,
388 &ExportLists);
389 }
390
391 #ifndef NDEBUG
392 DEBUG(dbgs() << "Import/Export lists for " << ImportLists.size()
393 << " modules:\n");
394 for (auto &ModuleImports : ImportLists) {
395 auto ModName = ModuleImports.first();
396 auto &Exports = ExportLists[ModName];
397 DEBUG(dbgs() << "* Module " << ModName << " exports " << Exports.size()
398 << " functions. Imports from " << ModuleImports.second.size()
399 << " modules.\n");
400 for (auto &Src : ModuleImports.second) {
401 auto SrcModName = Src.first();
402 DEBUG(dbgs() << " - " << Src.second.size() << " functions imported from "
403 << SrcModName << "\n");
404 }
405 }
406 #endif
407 }
408
409 /// Compute all the imports for the given module in the Index.
ComputeCrossModuleImportForModule(StringRef ModulePath,const ModuleSummaryIndex & Index,FunctionImporter::ImportMapTy & ImportList)410 void llvm::ComputeCrossModuleImportForModule(
411 StringRef ModulePath, const ModuleSummaryIndex &Index,
412 FunctionImporter::ImportMapTy &ImportList) {
413
414 // Collect the list of functions this module defines.
415 // GUID -> Summary
416 GVSummaryMapTy FunctionSummaryMap;
417 Index.collectDefinedFunctionsForModule(ModulePath, FunctionSummaryMap);
418
419 // Compute the import list for this module.
420 DEBUG(dbgs() << "Computing import for Module '" << ModulePath << "'\n");
421 ComputeImportForModule(FunctionSummaryMap, Index, ImportList);
422
423 #ifndef NDEBUG
424 DEBUG(dbgs() << "* Module " << ModulePath << " imports from "
425 << ImportList.size() << " modules.\n");
426 for (auto &Src : ImportList) {
427 auto SrcModName = Src.first();
428 DEBUG(dbgs() << " - " << Src.second.size() << " functions imported from "
429 << SrcModName << "\n");
430 }
431 #endif
432 }
433
434 /// Compute the set of summaries needed for a ThinLTO backend compilation of
435 /// \p ModulePath.
gatherImportedSummariesForModule(StringRef ModulePath,const StringMap<GVSummaryMapTy> & ModuleToDefinedGVSummaries,const StringMap<FunctionImporter::ImportMapTy> & ImportLists,std::map<std::string,GVSummaryMapTy> & ModuleToSummariesForIndex)436 void llvm::gatherImportedSummariesForModule(
437 StringRef ModulePath,
438 const StringMap<GVSummaryMapTy> &ModuleToDefinedGVSummaries,
439 const StringMap<FunctionImporter::ImportMapTy> &ImportLists,
440 std::map<std::string, GVSummaryMapTy> &ModuleToSummariesForIndex) {
441 // Include all summaries from the importing module.
442 ModuleToSummariesForIndex[ModulePath] =
443 ModuleToDefinedGVSummaries.lookup(ModulePath);
444 auto ModuleImports = ImportLists.find(ModulePath);
445 if (ModuleImports != ImportLists.end()) {
446 // Include summaries for imports.
447 for (auto &ILI : ModuleImports->second) {
448 auto &SummariesForIndex = ModuleToSummariesForIndex[ILI.first()];
449 const auto &DefinedGVSummaries =
450 ModuleToDefinedGVSummaries.lookup(ILI.first());
451 for (auto &GI : ILI.second) {
452 const auto &DS = DefinedGVSummaries.find(GI.first);
453 assert(DS != DefinedGVSummaries.end() &&
454 "Expected a defined summary for imported global value");
455 SummariesForIndex[GI.first] = DS->second;
456 }
457 }
458 }
459 }
460
461 /// Emit the files \p ModulePath will import from into \p OutputFilename.
EmitImportsFiles(StringRef ModulePath,StringRef OutputFilename,const StringMap<FunctionImporter::ImportMapTy> & ImportLists)462 std::error_code llvm::EmitImportsFiles(
463 StringRef ModulePath, StringRef OutputFilename,
464 const StringMap<FunctionImporter::ImportMapTy> &ImportLists) {
465 auto ModuleImports = ImportLists.find(ModulePath);
466 std::error_code EC;
467 raw_fd_ostream ImportsOS(OutputFilename, EC, sys::fs::OpenFlags::F_None);
468 if (EC)
469 return EC;
470 if (ModuleImports != ImportLists.end())
471 for (auto &ILI : ModuleImports->second)
472 ImportsOS << ILI.first() << "\n";
473 return std::error_code();
474 }
475
476 /// Fixup WeakForLinker linkages in \p TheModule based on summary analysis.
thinLTOResolveWeakForLinkerModule(Module & TheModule,const GVSummaryMapTy & DefinedGlobals)477 void llvm::thinLTOResolveWeakForLinkerModule(
478 Module &TheModule, const GVSummaryMapTy &DefinedGlobals) {
479 auto updateLinkage = [&](GlobalValue &GV) {
480 if (!GlobalValue::isWeakForLinker(GV.getLinkage()))
481 return;
482 // See if the global summary analysis computed a new resolved linkage.
483 const auto &GS = DefinedGlobals.find(GV.getGUID());
484 if (GS == DefinedGlobals.end())
485 return;
486 auto NewLinkage = GS->second->linkage();
487 if (NewLinkage == GV.getLinkage())
488 return;
489 DEBUG(dbgs() << "ODR fixing up linkage for `" << GV.getName() << "` from "
490 << GV.getLinkage() << " to " << NewLinkage << "\n");
491 GV.setLinkage(NewLinkage);
492 };
493
494 // Process functions and global now
495 for (auto &GV : TheModule)
496 updateLinkage(GV);
497 for (auto &GV : TheModule.globals())
498 updateLinkage(GV);
499 for (auto &GV : TheModule.aliases())
500 updateLinkage(GV);
501 }
502
503 /// Run internalization on \p TheModule based on symmary analysis.
thinLTOInternalizeModule(Module & TheModule,const GVSummaryMapTy & DefinedGlobals)504 void llvm::thinLTOInternalizeModule(Module &TheModule,
505 const GVSummaryMapTy &DefinedGlobals) {
506 // Parse inline ASM and collect the list of symbols that are not defined in
507 // the current module.
508 StringSet<> AsmUndefinedRefs;
509 object::IRObjectFile::CollectAsmUndefinedRefs(
510 Triple(TheModule.getTargetTriple()), TheModule.getModuleInlineAsm(),
511 [&AsmUndefinedRefs](StringRef Name, object::BasicSymbolRef::Flags Flags) {
512 if (Flags & object::BasicSymbolRef::SF_Undefined)
513 AsmUndefinedRefs.insert(Name);
514 });
515
516 // Declare a callback for the internalize pass that will ask for every
517 // candidate GlobalValue if it can be internalized or not.
518 auto MustPreserveGV = [&](const GlobalValue &GV) -> bool {
519 // Can't be internalized if referenced in inline asm.
520 if (AsmUndefinedRefs.count(GV.getName()))
521 return true;
522
523 // Lookup the linkage recorded in the summaries during global analysis.
524 const auto &GS = DefinedGlobals.find(GV.getGUID());
525 GlobalValue::LinkageTypes Linkage;
526 if (GS == DefinedGlobals.end()) {
527 // Must have been promoted (possibly conservatively). Find original
528 // name so that we can access the correct summary and see if it can
529 // be internalized again.
530 // FIXME: Eventually we should control promotion instead of promoting
531 // and internalizing again.
532 StringRef OrigName =
533 ModuleSummaryIndex::getOriginalNameBeforePromote(GV.getName());
534 std::string OrigId = GlobalValue::getGlobalIdentifier(
535 OrigName, GlobalValue::InternalLinkage,
536 TheModule.getSourceFileName());
537 const auto &GS = DefinedGlobals.find(GlobalValue::getGUID(OrigId));
538 if (GS == DefinedGlobals.end()) {
539 // Also check the original non-promoted non-globalized name. In some
540 // cases a preempted weak value is linked in as a local copy because
541 // it is referenced by an alias (IRLinker::linkGlobalValueProto).
542 // In that case, since it was originally not a local value, it was
543 // recorded in the index using the original name.
544 // FIXME: This may not be needed once PR27866 is fixed.
545 const auto &GS = DefinedGlobals.find(GlobalValue::getGUID(OrigName));
546 assert(GS != DefinedGlobals.end());
547 Linkage = GS->second->linkage();
548 } else {
549 Linkage = GS->second->linkage();
550 }
551 } else
552 Linkage = GS->second->linkage();
553 return !GlobalValue::isLocalLinkage(Linkage);
554 };
555
556 // FIXME: See if we can just internalize directly here via linkage changes
557 // based on the index, rather than invoking internalizeModule.
558 llvm::internalizeModule(TheModule, MustPreserveGV);
559 }
560
561 // Automatically import functions in Module \p DestModule based on the summaries
562 // index.
563 //
importFunctions(Module & DestModule,const FunctionImporter::ImportMapTy & ImportList,bool ForceImportReferencedDiscardableSymbols)564 bool FunctionImporter::importFunctions(
565 Module &DestModule, const FunctionImporter::ImportMapTy &ImportList,
566 bool ForceImportReferencedDiscardableSymbols) {
567 DEBUG(dbgs() << "Starting import for Module "
568 << DestModule.getModuleIdentifier() << "\n");
569 unsigned ImportedCount = 0;
570
571 // Linker that will be used for importing function
572 Linker TheLinker(DestModule);
573 // Do the actual import of functions now, one Module at a time
574 std::set<StringRef> ModuleNameOrderedList;
575 for (auto &FunctionsToImportPerModule : ImportList) {
576 ModuleNameOrderedList.insert(FunctionsToImportPerModule.first());
577 }
578 for (auto &Name : ModuleNameOrderedList) {
579 // Get the module for the import
580 const auto &FunctionsToImportPerModule = ImportList.find(Name);
581 assert(FunctionsToImportPerModule != ImportList.end());
582 std::unique_ptr<Module> SrcModule = ModuleLoader(Name);
583 assert(&DestModule.getContext() == &SrcModule->getContext() &&
584 "Context mismatch");
585
586 // If modules were created with lazy metadata loading, materialize it
587 // now, before linking it (otherwise this will be a noop).
588 SrcModule->materializeMetadata();
589 UpgradeDebugInfo(*SrcModule);
590
591 auto &ImportGUIDs = FunctionsToImportPerModule->second;
592 // Find the globals to import
593 DenseSet<const GlobalValue *> GlobalsToImport;
594 for (Function &F : *SrcModule) {
595 if (!F.hasName())
596 continue;
597 auto GUID = F.getGUID();
598 auto Import = ImportGUIDs.count(GUID);
599 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing function " << GUID
600 << " " << F.getName() << " from "
601 << SrcModule->getSourceFileName() << "\n");
602 if (Import) {
603 F.materialize();
604 if (EnableImportMetadata) {
605 // Add 'thinlto_src_module' metadata for statistics and debugging.
606 F.setMetadata(
607 "thinlto_src_module",
608 llvm::MDNode::get(
609 DestModule.getContext(),
610 {llvm::MDString::get(DestModule.getContext(),
611 SrcModule->getSourceFileName())}));
612 }
613 GlobalsToImport.insert(&F);
614 }
615 }
616 for (GlobalVariable &GV : SrcModule->globals()) {
617 if (!GV.hasName())
618 continue;
619 auto GUID = GV.getGUID();
620 auto Import = ImportGUIDs.count(GUID);
621 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing global " << GUID
622 << " " << GV.getName() << " from "
623 << SrcModule->getSourceFileName() << "\n");
624 if (Import) {
625 GV.materialize();
626 GlobalsToImport.insert(&GV);
627 }
628 }
629 for (GlobalAlias &GA : SrcModule->aliases()) {
630 if (!GA.hasName())
631 continue;
632 auto GUID = GA.getGUID();
633 auto Import = ImportGUIDs.count(GUID);
634 DEBUG(dbgs() << (Import ? "Is" : "Not") << " importing alias " << GUID
635 << " " << GA.getName() << " from "
636 << SrcModule->getSourceFileName() << "\n");
637 if (Import) {
638 // Alias can't point to "available_externally". However when we import
639 // linkOnceODR the linkage does not change. So we import the alias
640 // and aliasee only in this case. This has been handled by
641 // computeImportForFunction()
642 GlobalObject *GO = GA.getBaseObject();
643 assert(GO->hasLinkOnceODRLinkage() &&
644 "Unexpected alias to a non-linkonceODR in import list");
645 #ifndef NDEBUG
646 if (!GlobalsToImport.count(GO))
647 DEBUG(dbgs() << " alias triggers importing aliasee " << GO->getGUID()
648 << " " << GO->getName() << " from "
649 << SrcModule->getSourceFileName() << "\n");
650 #endif
651 GO->materialize();
652 GlobalsToImport.insert(GO);
653 GA.materialize();
654 GlobalsToImport.insert(&GA);
655 }
656 }
657
658 // Link in the specified functions.
659 if (renameModuleForThinLTO(*SrcModule, Index, &GlobalsToImport))
660 return true;
661
662 if (PrintImports) {
663 for (const auto *GV : GlobalsToImport)
664 dbgs() << DestModule.getSourceFileName() << ": Import " << GV->getName()
665 << " from " << SrcModule->getSourceFileName() << "\n";
666 }
667
668 // Instruct the linker that the client will take care of linkonce resolution
669 unsigned Flags = Linker::Flags::None;
670 if (!ForceImportReferencedDiscardableSymbols)
671 Flags |= Linker::Flags::DontForceLinkLinkonceODR;
672
673 if (TheLinker.linkInModule(std::move(SrcModule), Flags, &GlobalsToImport))
674 report_fatal_error("Function Import: link error");
675
676 ImportedCount += GlobalsToImport.size();
677 }
678
679 NumImported += ImportedCount;
680
681 DEBUG(dbgs() << "Imported " << ImportedCount << " functions for Module "
682 << DestModule.getModuleIdentifier() << "\n");
683 return ImportedCount;
684 }
685
686 /// Summary file to use for function importing when using -function-import from
687 /// the command line.
688 static cl::opt<std::string>
689 SummaryFile("summary-file",
690 cl::desc("The summary file to use for function importing."));
691
diagnosticHandler(const DiagnosticInfo & DI)692 static void diagnosticHandler(const DiagnosticInfo &DI) {
693 raw_ostream &OS = errs();
694 DiagnosticPrinterRawOStream DP(OS);
695 DI.print(DP);
696 OS << '\n';
697 }
698
699 /// Parse the summary index out of an IR file and return the summary
700 /// index object if found, or nullptr if not.
getModuleSummaryIndexForFile(StringRef Path,std::string & Error,const DiagnosticHandlerFunction & DiagnosticHandler)701 static std::unique_ptr<ModuleSummaryIndex> getModuleSummaryIndexForFile(
702 StringRef Path, std::string &Error,
703 const DiagnosticHandlerFunction &DiagnosticHandler) {
704 std::unique_ptr<MemoryBuffer> Buffer;
705 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
706 MemoryBuffer::getFile(Path);
707 if (std::error_code EC = BufferOrErr.getError()) {
708 Error = EC.message();
709 return nullptr;
710 }
711 Buffer = std::move(BufferOrErr.get());
712 ErrorOr<std::unique_ptr<object::ModuleSummaryIndexObjectFile>> ObjOrErr =
713 object::ModuleSummaryIndexObjectFile::create(Buffer->getMemBufferRef(),
714 DiagnosticHandler);
715 if (std::error_code EC = ObjOrErr.getError()) {
716 Error = EC.message();
717 return nullptr;
718 }
719 return (*ObjOrErr)->takeIndex();
720 }
721
722 namespace {
723 /// Pass that performs cross-module function import provided a summary file.
724 class FunctionImportPass : public ModulePass {
725 /// Optional module summary index to use for importing, otherwise
726 /// the summary-file option must be specified.
727 const ModuleSummaryIndex *Index;
728
729 public:
730 /// Pass identification, replacement for typeid
731 static char ID;
732
733 /// Specify pass name for debug output
getPassName() const734 const char *getPassName() const override { return "Function Importing"; }
735
FunctionImportPass(const ModuleSummaryIndex * Index=nullptr)736 explicit FunctionImportPass(const ModuleSummaryIndex *Index = nullptr)
737 : ModulePass(ID), Index(Index) {}
738
runOnModule(Module & M)739 bool runOnModule(Module &M) override {
740 if (skipModule(M))
741 return false;
742
743 if (SummaryFile.empty() && !Index)
744 report_fatal_error("error: -function-import requires -summary-file or "
745 "file from frontend\n");
746 std::unique_ptr<ModuleSummaryIndex> IndexPtr;
747 if (!SummaryFile.empty()) {
748 if (Index)
749 report_fatal_error("error: -summary-file and index from frontend\n");
750 std::string Error;
751 IndexPtr =
752 getModuleSummaryIndexForFile(SummaryFile, Error, diagnosticHandler);
753 if (!IndexPtr) {
754 errs() << "Error loading file '" << SummaryFile << "': " << Error
755 << "\n";
756 return false;
757 }
758 Index = IndexPtr.get();
759 }
760
761 // First step is collecting the import list.
762 FunctionImporter::ImportMapTy ImportList;
763 ComputeCrossModuleImportForModule(M.getModuleIdentifier(), *Index,
764 ImportList);
765
766 // Next we need to promote to global scope and rename any local values that
767 // are potentially exported to other modules.
768 if (renameModuleForThinLTO(M, *Index, nullptr)) {
769 errs() << "Error renaming module\n";
770 return false;
771 }
772
773 // Perform the import now.
774 auto ModuleLoader = [&M](StringRef Identifier) {
775 return loadFile(Identifier, M.getContext());
776 };
777 FunctionImporter Importer(*Index, ModuleLoader);
778 return Importer.importFunctions(
779 M, ImportList, !DontForceImportReferencedDiscardableSymbols);
780 }
781 };
782 } // anonymous namespace
783
784 char FunctionImportPass::ID = 0;
785 INITIALIZE_PASS_BEGIN(FunctionImportPass, "function-import",
786 "Summary Based Function Import", false, false)
787 INITIALIZE_PASS_END(FunctionImportPass, "function-import",
788 "Summary Based Function Import", false, false)
789
790 namespace llvm {
createFunctionImportPass(const ModuleSummaryIndex * Index=nullptr)791 Pass *createFunctionImportPass(const ModuleSummaryIndex *Index = nullptr) {
792 return new FunctionImportPass(Index);
793 }
794 }
795