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