1 //===- ModuleSummaryAnalysis.cpp - Module summary index builder -----------===// 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 pass builds a ModuleSummaryIndex object for the module, to be written 10 // to bitcode or LLVM assembly. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Analysis/ModuleSummaryAnalysis.h" 15 #include "llvm/ADT/ArrayRef.h" 16 #include "llvm/ADT/DenseSet.h" 17 #include "llvm/ADT/MapVector.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SetVector.h" 20 #include "llvm/ADT/SmallPtrSet.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/ADT/StringRef.h" 23 #include "llvm/Analysis/BlockFrequencyInfo.h" 24 #include "llvm/Analysis/BranchProbabilityInfo.h" 25 #include "llvm/Analysis/ConstantFolding.h" 26 #include "llvm/Analysis/IndirectCallPromotionAnalysis.h" 27 #include "llvm/Analysis/LoopInfo.h" 28 #include "llvm/Analysis/MemoryProfileInfo.h" 29 #include "llvm/Analysis/ProfileSummaryInfo.h" 30 #include "llvm/Analysis/StackSafetyAnalysis.h" 31 #include "llvm/Analysis/TypeMetadataUtils.h" 32 #include "llvm/IR/Attributes.h" 33 #include "llvm/IR/BasicBlock.h" 34 #include "llvm/IR/Constant.h" 35 #include "llvm/IR/Constants.h" 36 #include "llvm/IR/Dominators.h" 37 #include "llvm/IR/Function.h" 38 #include "llvm/IR/GlobalAlias.h" 39 #include "llvm/IR/GlobalValue.h" 40 #include "llvm/IR/GlobalVariable.h" 41 #include "llvm/IR/Instructions.h" 42 #include "llvm/IR/IntrinsicInst.h" 43 #include "llvm/IR/Metadata.h" 44 #include "llvm/IR/Module.h" 45 #include "llvm/IR/ModuleSummaryIndex.h" 46 #include "llvm/IR/Use.h" 47 #include "llvm/IR/User.h" 48 #include "llvm/InitializePasses.h" 49 #include "llvm/Object/ModuleSymbolTable.h" 50 #include "llvm/Object/SymbolicFile.h" 51 #include "llvm/Pass.h" 52 #include "llvm/Support/Casting.h" 53 #include "llvm/Support/CommandLine.h" 54 #include "llvm/Support/FileSystem.h" 55 #include <algorithm> 56 #include <cassert> 57 #include <cstdint> 58 #include <vector> 59 60 using namespace llvm; 61 using namespace llvm::memprof; 62 63 #define DEBUG_TYPE "module-summary-analysis" 64 65 // Option to force edges cold which will block importing when the 66 // -import-cold-multiplier is set to 0. Useful for debugging. 67 namespace llvm { 68 FunctionSummary::ForceSummaryHotnessType ForceSummaryEdgesCold = 69 FunctionSummary::FSHT_None; 70 } // namespace llvm 71 72 static cl::opt<FunctionSummary::ForceSummaryHotnessType, true> FSEC( 73 "force-summary-edges-cold", cl::Hidden, cl::location(ForceSummaryEdgesCold), 74 cl::desc("Force all edges in the function summary to cold"), 75 cl::values(clEnumValN(FunctionSummary::FSHT_None, "none", "None."), 76 clEnumValN(FunctionSummary::FSHT_AllNonCritical, 77 "all-non-critical", "All non-critical edges."), 78 clEnumValN(FunctionSummary::FSHT_All, "all", "All edges."))); 79 80 static cl::opt<std::string> ModuleSummaryDotFile( 81 "module-summary-dot-file", cl::Hidden, cl::value_desc("filename"), 82 cl::desc("File to emit dot graph of new summary into")); 83 84 extern cl::opt<bool> ScalePartialSampleProfileWorkingSetSize; 85 86 extern cl::opt<unsigned> MaxNumVTableAnnotations; 87 88 // Walk through the operands of a given User via worklist iteration and populate 89 // the set of GlobalValue references encountered. Invoked either on an 90 // Instruction or a GlobalVariable (which walks its initializer). 91 // Return true if any of the operands contains blockaddress. This is important 92 // to know when computing summary for global var, because if global variable 93 // references basic block address we can't import it separately from function 94 // containing that basic block. For simplicity we currently don't import such 95 // global vars at all. When importing function we aren't interested if any 96 // instruction in it takes an address of any basic block, because instruction 97 // can only take an address of basic block located in the same function. 98 // Set `RefLocalLinkageIFunc` to true if the analyzed value references a 99 // local-linkage ifunc. 100 static bool findRefEdges(ModuleSummaryIndex &Index, const User *CurUser, 101 SetVector<ValueInfo, std::vector<ValueInfo>> &RefEdges, 102 SmallPtrSet<const User *, 8> &Visited, 103 bool &RefLocalLinkageIFunc) { 104 bool HasBlockAddress = false; 105 SmallVector<const User *, 32> Worklist; 106 if (Visited.insert(CurUser).second) 107 Worklist.push_back(CurUser); 108 109 while (!Worklist.empty()) { 110 const User *U = Worklist.pop_back_val(); 111 const auto *CB = dyn_cast<CallBase>(U); 112 113 for (const auto &OI : U->operands()) { 114 const User *Operand = dyn_cast<User>(OI); 115 if (!Operand) 116 continue; 117 if (isa<BlockAddress>(Operand)) { 118 HasBlockAddress = true; 119 continue; 120 } 121 if (auto *GV = dyn_cast<GlobalValue>(Operand)) { 122 // We have a reference to a global value. This should be added to 123 // the reference set unless it is a callee. Callees are handled 124 // specially by WriteFunction and are added to a separate list. 125 if (!(CB && CB->isCallee(&OI))) { 126 // If an ifunc has local linkage, do not add it into ref edges, and 127 // sets `RefLocalLinkageIFunc` to true. The referencer is not eligible 128 // for import. An ifunc doesn't have summary and ThinLTO cannot 129 // promote it; importing the referencer may cause linkage errors. 130 if (auto *GI = dyn_cast_if_present<GlobalIFunc>(GV); 131 GI && GI->hasLocalLinkage()) { 132 RefLocalLinkageIFunc = true; 133 continue; 134 } 135 RefEdges.insert(Index.getOrInsertValueInfo(GV)); 136 } 137 continue; 138 } 139 if (Visited.insert(Operand).second) 140 Worklist.push_back(Operand); 141 } 142 } 143 144 const Instruction *I = dyn_cast<Instruction>(CurUser); 145 if (I) { 146 uint32_t ActualNumValueData = 0; 147 uint64_t TotalCount = 0; 148 // MaxNumVTableAnnotations is the maximum number of vtables annotated on 149 // the instruction. 150 auto ValueDataArray = 151 getValueProfDataFromInst(*I, IPVK_VTableTarget, MaxNumVTableAnnotations, 152 ActualNumValueData, TotalCount); 153 154 if (ValueDataArray.get()) { 155 for (uint32_t j = 0; j < ActualNumValueData; j++) { 156 RefEdges.insert(Index.getOrInsertValueInfo(/* VTableGUID = */ 157 ValueDataArray[j].Value)); 158 } 159 } 160 } 161 return HasBlockAddress; 162 } 163 164 static CalleeInfo::HotnessType getHotness(uint64_t ProfileCount, 165 ProfileSummaryInfo *PSI) { 166 if (!PSI) 167 return CalleeInfo::HotnessType::Unknown; 168 if (PSI->isHotCount(ProfileCount)) 169 return CalleeInfo::HotnessType::Hot; 170 if (PSI->isColdCount(ProfileCount)) 171 return CalleeInfo::HotnessType::Cold; 172 return CalleeInfo::HotnessType::None; 173 } 174 175 static bool isNonRenamableLocal(const GlobalValue &GV) { 176 return GV.hasSection() && GV.hasLocalLinkage(); 177 } 178 179 /// Determine whether this call has all constant integer arguments (excluding 180 /// "this") and summarize it to VCalls or ConstVCalls as appropriate. 181 static void addVCallToSet( 182 DevirtCallSite Call, GlobalValue::GUID Guid, 183 SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>> 184 &VCalls, 185 SetVector<FunctionSummary::ConstVCall, 186 std::vector<FunctionSummary::ConstVCall>> &ConstVCalls) { 187 std::vector<uint64_t> Args; 188 // Start from the second argument to skip the "this" pointer. 189 for (auto &Arg : drop_begin(Call.CB.args())) { 190 auto *CI = dyn_cast<ConstantInt>(Arg); 191 if (!CI || CI->getBitWidth() > 64) { 192 VCalls.insert({Guid, Call.Offset}); 193 return; 194 } 195 Args.push_back(CI->getZExtValue()); 196 } 197 ConstVCalls.insert({{Guid, Call.Offset}, std::move(Args)}); 198 } 199 200 /// If this intrinsic call requires that we add information to the function 201 /// summary, do so via the non-constant reference arguments. 202 static void addIntrinsicToSummary( 203 const CallInst *CI, 204 SetVector<GlobalValue::GUID, std::vector<GlobalValue::GUID>> &TypeTests, 205 SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>> 206 &TypeTestAssumeVCalls, 207 SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>> 208 &TypeCheckedLoadVCalls, 209 SetVector<FunctionSummary::ConstVCall, 210 std::vector<FunctionSummary::ConstVCall>> 211 &TypeTestAssumeConstVCalls, 212 SetVector<FunctionSummary::ConstVCall, 213 std::vector<FunctionSummary::ConstVCall>> 214 &TypeCheckedLoadConstVCalls, 215 DominatorTree &DT) { 216 switch (CI->getCalledFunction()->getIntrinsicID()) { 217 case Intrinsic::type_test: 218 case Intrinsic::public_type_test: { 219 auto *TypeMDVal = cast<MetadataAsValue>(CI->getArgOperand(1)); 220 auto *TypeId = dyn_cast<MDString>(TypeMDVal->getMetadata()); 221 if (!TypeId) 222 break; 223 GlobalValue::GUID Guid = GlobalValue::getGUID(TypeId->getString()); 224 225 // Produce a summary from type.test intrinsics. We only summarize type.test 226 // intrinsics that are used other than by an llvm.assume intrinsic. 227 // Intrinsics that are assumed are relevant only to the devirtualization 228 // pass, not the type test lowering pass. 229 bool HasNonAssumeUses = llvm::any_of(CI->uses(), [](const Use &CIU) { 230 return !isa<AssumeInst>(CIU.getUser()); 231 }); 232 if (HasNonAssumeUses) 233 TypeTests.insert(Guid); 234 235 SmallVector<DevirtCallSite, 4> DevirtCalls; 236 SmallVector<CallInst *, 4> Assumes; 237 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT); 238 for (auto &Call : DevirtCalls) 239 addVCallToSet(Call, Guid, TypeTestAssumeVCalls, 240 TypeTestAssumeConstVCalls); 241 242 break; 243 } 244 245 case Intrinsic::type_checked_load_relative: 246 case Intrinsic::type_checked_load: { 247 auto *TypeMDVal = cast<MetadataAsValue>(CI->getArgOperand(2)); 248 auto *TypeId = dyn_cast<MDString>(TypeMDVal->getMetadata()); 249 if (!TypeId) 250 break; 251 GlobalValue::GUID Guid = GlobalValue::getGUID(TypeId->getString()); 252 253 SmallVector<DevirtCallSite, 4> DevirtCalls; 254 SmallVector<Instruction *, 4> LoadedPtrs; 255 SmallVector<Instruction *, 4> Preds; 256 bool HasNonCallUses = false; 257 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds, 258 HasNonCallUses, CI, DT); 259 // Any non-call uses of the result of llvm.type.checked.load will 260 // prevent us from optimizing away the llvm.type.test. 261 if (HasNonCallUses) 262 TypeTests.insert(Guid); 263 for (auto &Call : DevirtCalls) 264 addVCallToSet(Call, Guid, TypeCheckedLoadVCalls, 265 TypeCheckedLoadConstVCalls); 266 267 break; 268 } 269 default: 270 break; 271 } 272 } 273 274 static bool isNonVolatileLoad(const Instruction *I) { 275 if (const auto *LI = dyn_cast<LoadInst>(I)) 276 return !LI->isVolatile(); 277 278 return false; 279 } 280 281 static bool isNonVolatileStore(const Instruction *I) { 282 if (const auto *SI = dyn_cast<StoreInst>(I)) 283 return !SI->isVolatile(); 284 285 return false; 286 } 287 288 // Returns true if the function definition must be unreachable. 289 // 290 // Note if this helper function returns true, `F` is guaranteed 291 // to be unreachable; if it returns false, `F` might still 292 // be unreachable but not covered by this helper function. 293 static bool mustBeUnreachableFunction(const Function &F) { 294 // A function must be unreachable if its entry block ends with an 295 // 'unreachable'. 296 assert(!F.isDeclaration()); 297 return isa<UnreachableInst>(F.getEntryBlock().getTerminator()); 298 } 299 300 static void computeFunctionSummary( 301 ModuleSummaryIndex &Index, const Module &M, const Function &F, 302 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, DominatorTree &DT, 303 bool HasLocalsInUsedOrAsm, DenseSet<GlobalValue::GUID> &CantBePromoted, 304 bool IsThinLTO, 305 std::function<const StackSafetyInfo *(const Function &F)> GetSSICallback) { 306 // Summary not currently supported for anonymous functions, they should 307 // have been named. 308 assert(F.hasName()); 309 310 unsigned NumInsts = 0; 311 // Map from callee ValueId to profile count. Used to accumulate profile 312 // counts for all static calls to a given callee. 313 MapVector<ValueInfo, CalleeInfo, DenseMap<ValueInfo, unsigned>, 314 std::vector<std::pair<ValueInfo, CalleeInfo>>> 315 CallGraphEdges; 316 SetVector<ValueInfo, std::vector<ValueInfo>> RefEdges, LoadRefEdges, 317 StoreRefEdges; 318 SetVector<GlobalValue::GUID, std::vector<GlobalValue::GUID>> TypeTests; 319 SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>> 320 TypeTestAssumeVCalls, TypeCheckedLoadVCalls; 321 SetVector<FunctionSummary::ConstVCall, 322 std::vector<FunctionSummary::ConstVCall>> 323 TypeTestAssumeConstVCalls, TypeCheckedLoadConstVCalls; 324 ICallPromotionAnalysis ICallAnalysis; 325 SmallPtrSet<const User *, 8> Visited; 326 327 // Add personality function, prefix data and prologue data to function's ref 328 // list. 329 bool HasLocalIFuncCallOrRef = false; 330 findRefEdges(Index, &F, RefEdges, Visited, HasLocalIFuncCallOrRef); 331 std::vector<const Instruction *> NonVolatileLoads; 332 std::vector<const Instruction *> NonVolatileStores; 333 334 std::vector<CallsiteInfo> Callsites; 335 std::vector<AllocInfo> Allocs; 336 337 #ifndef NDEBUG 338 DenseSet<const CallBase *> CallsThatMayHaveMemprofSummary; 339 #endif 340 341 bool HasInlineAsmMaybeReferencingInternal = false; 342 bool HasIndirBranchToBlockAddress = false; 343 bool HasUnknownCall = false; 344 bool MayThrow = false; 345 for (const BasicBlock &BB : F) { 346 // We don't allow inlining of function with indirect branch to blockaddress. 347 // If the blockaddress escapes the function, e.g., via a global variable, 348 // inlining may lead to an invalid cross-function reference. So we shouldn't 349 // import such function either. 350 if (BB.hasAddressTaken()) { 351 for (User *U : BlockAddress::get(const_cast<BasicBlock *>(&BB))->users()) 352 if (!isa<CallBrInst>(*U)) { 353 HasIndirBranchToBlockAddress = true; 354 break; 355 } 356 } 357 358 for (const Instruction &I : BB) { 359 if (I.isDebugOrPseudoInst()) 360 continue; 361 ++NumInsts; 362 363 // Regular LTO module doesn't participate in ThinLTO import, 364 // so no reference from it can be read/writeonly, since this 365 // would require importing variable as local copy 366 if (IsThinLTO) { 367 if (isNonVolatileLoad(&I)) { 368 // Postpone processing of non-volatile load instructions 369 // See comments below 370 Visited.insert(&I); 371 NonVolatileLoads.push_back(&I); 372 continue; 373 } else if (isNonVolatileStore(&I)) { 374 Visited.insert(&I); 375 NonVolatileStores.push_back(&I); 376 // All references from second operand of store (destination address) 377 // can be considered write-only if they're not referenced by any 378 // non-store instruction. References from first operand of store 379 // (stored value) can't be treated either as read- or as write-only 380 // so we add them to RefEdges as we do with all other instructions 381 // except non-volatile load. 382 Value *Stored = I.getOperand(0); 383 if (auto *GV = dyn_cast<GlobalValue>(Stored)) 384 // findRefEdges will try to examine GV operands, so instead 385 // of calling it we should add GV to RefEdges directly. 386 RefEdges.insert(Index.getOrInsertValueInfo(GV)); 387 else if (auto *U = dyn_cast<User>(Stored)) 388 findRefEdges(Index, U, RefEdges, Visited, HasLocalIFuncCallOrRef); 389 continue; 390 } 391 } 392 findRefEdges(Index, &I, RefEdges, Visited, HasLocalIFuncCallOrRef); 393 const auto *CB = dyn_cast<CallBase>(&I); 394 if (!CB) { 395 if (I.mayThrow()) 396 MayThrow = true; 397 continue; 398 } 399 400 const auto *CI = dyn_cast<CallInst>(&I); 401 // Since we don't know exactly which local values are referenced in inline 402 // assembly, conservatively mark the function as possibly referencing 403 // a local value from inline assembly to ensure we don't export a 404 // reference (which would require renaming and promotion of the 405 // referenced value). 406 if (HasLocalsInUsedOrAsm && CI && CI->isInlineAsm()) 407 HasInlineAsmMaybeReferencingInternal = true; 408 409 auto *CalledValue = CB->getCalledOperand(); 410 auto *CalledFunction = CB->getCalledFunction(); 411 if (CalledValue && !CalledFunction) { 412 CalledValue = CalledValue->stripPointerCasts(); 413 // Stripping pointer casts can reveal a called function. 414 CalledFunction = dyn_cast<Function>(CalledValue); 415 } 416 // Check if this is an alias to a function. If so, get the 417 // called aliasee for the checks below. 418 if (auto *GA = dyn_cast<GlobalAlias>(CalledValue)) { 419 assert(!CalledFunction && "Expected null called function in callsite for alias"); 420 CalledFunction = dyn_cast<Function>(GA->getAliaseeObject()); 421 } 422 // Check if this is a direct call to a known function or a known 423 // intrinsic, or an indirect call with profile data. 424 if (CalledFunction) { 425 if (CI && CalledFunction->isIntrinsic()) { 426 addIntrinsicToSummary( 427 CI, TypeTests, TypeTestAssumeVCalls, TypeCheckedLoadVCalls, 428 TypeTestAssumeConstVCalls, TypeCheckedLoadConstVCalls, DT); 429 continue; 430 } 431 // We should have named any anonymous globals 432 assert(CalledFunction->hasName()); 433 auto ScaledCount = PSI->getProfileCount(*CB, BFI); 434 auto Hotness = ScaledCount ? getHotness(*ScaledCount, PSI) 435 : CalleeInfo::HotnessType::Unknown; 436 if (ForceSummaryEdgesCold != FunctionSummary::FSHT_None) 437 Hotness = CalleeInfo::HotnessType::Cold; 438 439 // Use the original CalledValue, in case it was an alias. We want 440 // to record the call edge to the alias in that case. Eventually 441 // an alias summary will be created to associate the alias and 442 // aliasee. 443 auto &ValueInfo = CallGraphEdges[Index.getOrInsertValueInfo( 444 cast<GlobalValue>(CalledValue))]; 445 ValueInfo.updateHotness(Hotness); 446 if (CB->isTailCall()) 447 ValueInfo.setHasTailCall(true); 448 // Add the relative block frequency to CalleeInfo if there is no profile 449 // information. 450 if (BFI != nullptr && Hotness == CalleeInfo::HotnessType::Unknown) { 451 uint64_t BBFreq = BFI->getBlockFreq(&BB).getFrequency(); 452 uint64_t EntryFreq = BFI->getEntryFreq().getFrequency(); 453 ValueInfo.updateRelBlockFreq(BBFreq, EntryFreq); 454 } 455 } else { 456 HasUnknownCall = true; 457 // If F is imported, a local linkage ifunc (e.g. target_clones on a 458 // static function) called by F will be cloned. Since summaries don't 459 // track ifunc, we do not know implementation functions referenced by 460 // the ifunc resolver need to be promoted in the exporter, and we will 461 // get linker errors due to cloned declarations for implementation 462 // functions. As a simple fix, just mark F as not eligible for import. 463 // Non-local ifunc is not cloned and does not have the issue. 464 if (auto *GI = dyn_cast_if_present<GlobalIFunc>(CalledValue)) 465 if (GI->hasLocalLinkage()) 466 HasLocalIFuncCallOrRef = true; 467 // Skip inline assembly calls. 468 if (CI && CI->isInlineAsm()) 469 continue; 470 // Skip direct calls. 471 if (!CalledValue || isa<Constant>(CalledValue)) 472 continue; 473 474 // Check if the instruction has a callees metadata. If so, add callees 475 // to CallGraphEdges to reflect the references from the metadata, and 476 // to enable importing for subsequent indirect call promotion and 477 // inlining. 478 if (auto *MD = I.getMetadata(LLVMContext::MD_callees)) { 479 for (const auto &Op : MD->operands()) { 480 Function *Callee = mdconst::extract_or_null<Function>(Op); 481 if (Callee) 482 CallGraphEdges[Index.getOrInsertValueInfo(Callee)]; 483 } 484 } 485 486 uint32_t NumVals, NumCandidates; 487 uint64_t TotalCount; 488 auto CandidateProfileData = 489 ICallAnalysis.getPromotionCandidatesForInstruction( 490 &I, NumVals, TotalCount, NumCandidates); 491 for (const auto &Candidate : CandidateProfileData) 492 CallGraphEdges[Index.getOrInsertValueInfo(Candidate.Value)] 493 .updateHotness(getHotness(Candidate.Count, PSI)); 494 } 495 496 // Summarize memprof related metadata. This is only needed for ThinLTO. 497 if (!IsThinLTO) 498 continue; 499 500 // TODO: Skip indirect calls for now. Need to handle these better, likely 501 // by creating multiple Callsites, one per target, then speculatively 502 // devirtualize while applying clone info in the ThinLTO backends. This 503 // will also be important because we will have a different set of clone 504 // versions per target. This handling needs to match that in the ThinLTO 505 // backend so we handle things consistently for matching of callsite 506 // summaries to instructions. 507 if (!CalledFunction) 508 continue; 509 510 // Ensure we keep this analysis in sync with the handling in the ThinLTO 511 // backend (see MemProfContextDisambiguation::applyImport). Save this call 512 // so that we can skip it in checking the reverse case later. 513 assert(mayHaveMemprofSummary(CB)); 514 #ifndef NDEBUG 515 CallsThatMayHaveMemprofSummary.insert(CB); 516 #endif 517 518 // Compute the list of stack ids first (so we can trim them from the stack 519 // ids on any MIBs). 520 CallStack<MDNode, MDNode::op_iterator> InstCallsite( 521 I.getMetadata(LLVMContext::MD_callsite)); 522 auto *MemProfMD = I.getMetadata(LLVMContext::MD_memprof); 523 if (MemProfMD) { 524 std::vector<MIBInfo> MIBs; 525 for (auto &MDOp : MemProfMD->operands()) { 526 auto *MIBMD = cast<const MDNode>(MDOp); 527 MDNode *StackNode = getMIBStackNode(MIBMD); 528 assert(StackNode); 529 SmallVector<unsigned> StackIdIndices; 530 CallStack<MDNode, MDNode::op_iterator> StackContext(StackNode); 531 // Collapse out any on the allocation call (inlining). 532 for (auto ContextIter = 533 StackContext.beginAfterSharedPrefix(InstCallsite); 534 ContextIter != StackContext.end(); ++ContextIter) { 535 unsigned StackIdIdx = Index.addOrGetStackIdIndex(*ContextIter); 536 // If this is a direct recursion, simply skip the duplicate 537 // entries. If this is mutual recursion, handling is left to 538 // the LTO link analysis client. 539 if (StackIdIndices.empty() || StackIdIndices.back() != StackIdIdx) 540 StackIdIndices.push_back(StackIdIdx); 541 } 542 MIBs.push_back( 543 MIBInfo(getMIBAllocType(MIBMD), std::move(StackIdIndices))); 544 } 545 Allocs.push_back(AllocInfo(std::move(MIBs))); 546 } else if (!InstCallsite.empty()) { 547 SmallVector<unsigned> StackIdIndices; 548 for (auto StackId : InstCallsite) 549 StackIdIndices.push_back(Index.addOrGetStackIdIndex(StackId)); 550 // Use the original CalledValue, in case it was an alias. We want 551 // to record the call edge to the alias in that case. Eventually 552 // an alias summary will be created to associate the alias and 553 // aliasee. 554 auto CalleeValueInfo = 555 Index.getOrInsertValueInfo(cast<GlobalValue>(CalledValue)); 556 Callsites.push_back({CalleeValueInfo, StackIdIndices}); 557 } 558 } 559 } 560 561 if (PSI->hasPartialSampleProfile() && ScalePartialSampleProfileWorkingSetSize) 562 Index.addBlockCount(F.size()); 563 564 std::vector<ValueInfo> Refs; 565 if (IsThinLTO) { 566 auto AddRefEdges = [&](const std::vector<const Instruction *> &Instrs, 567 SetVector<ValueInfo, std::vector<ValueInfo>> &Edges, 568 SmallPtrSet<const User *, 8> &Cache) { 569 for (const auto *I : Instrs) { 570 Cache.erase(I); 571 findRefEdges(Index, I, Edges, Cache, HasLocalIFuncCallOrRef); 572 } 573 }; 574 575 // By now we processed all instructions in a function, except 576 // non-volatile loads and non-volatile value stores. Let's find 577 // ref edges for both of instruction sets 578 AddRefEdges(NonVolatileLoads, LoadRefEdges, Visited); 579 // We can add some values to the Visited set when processing load 580 // instructions which are also used by stores in NonVolatileStores. 581 // For example this can happen if we have following code: 582 // 583 // store %Derived* @foo, %Derived** bitcast (%Base** @bar to %Derived**) 584 // %42 = load %Derived*, %Derived** bitcast (%Base** @bar to %Derived**) 585 // 586 // After processing loads we'll add bitcast to the Visited set, and if 587 // we use the same set while processing stores, we'll never see store 588 // to @bar and @bar will be mistakenly treated as readonly. 589 SmallPtrSet<const llvm::User *, 8> StoreCache; 590 AddRefEdges(NonVolatileStores, StoreRefEdges, StoreCache); 591 592 // If both load and store instruction reference the same variable 593 // we won't be able to optimize it. Add all such reference edges 594 // to RefEdges set. 595 for (const auto &VI : StoreRefEdges) 596 if (LoadRefEdges.remove(VI)) 597 RefEdges.insert(VI); 598 599 unsigned RefCnt = RefEdges.size(); 600 // All new reference edges inserted in two loops below are either 601 // read or write only. They will be grouped in the end of RefEdges 602 // vector, so we can use a single integer value to identify them. 603 for (const auto &VI : LoadRefEdges) 604 RefEdges.insert(VI); 605 606 unsigned FirstWORef = RefEdges.size(); 607 for (const auto &VI : StoreRefEdges) 608 RefEdges.insert(VI); 609 610 Refs = RefEdges.takeVector(); 611 for (; RefCnt < FirstWORef; ++RefCnt) 612 Refs[RefCnt].setReadOnly(); 613 614 for (; RefCnt < Refs.size(); ++RefCnt) 615 Refs[RefCnt].setWriteOnly(); 616 } else { 617 Refs = RefEdges.takeVector(); 618 } 619 // Explicit add hot edges to enforce importing for designated GUIDs for 620 // sample PGO, to enable the same inlines as the profiled optimized binary. 621 for (auto &I : F.getImportGUIDs()) 622 CallGraphEdges[Index.getOrInsertValueInfo(I)].updateHotness( 623 ForceSummaryEdgesCold == FunctionSummary::FSHT_All 624 ? CalleeInfo::HotnessType::Cold 625 : CalleeInfo::HotnessType::Critical); 626 627 #ifndef NDEBUG 628 // Make sure that all calls we decided could not have memprof summaries get a 629 // false value for mayHaveMemprofSummary, to ensure that this handling remains 630 // in sync with the ThinLTO backend handling. 631 if (IsThinLTO) { 632 for (const BasicBlock &BB : F) { 633 for (const Instruction &I : BB) { 634 const auto *CB = dyn_cast<CallBase>(&I); 635 if (!CB) 636 continue; 637 // We already checked these above. 638 if (CallsThatMayHaveMemprofSummary.count(CB)) 639 continue; 640 assert(!mayHaveMemprofSummary(CB)); 641 } 642 } 643 } 644 #endif 645 646 bool NonRenamableLocal = isNonRenamableLocal(F); 647 bool NotEligibleForImport = 648 NonRenamableLocal || HasInlineAsmMaybeReferencingInternal || 649 HasIndirBranchToBlockAddress || HasLocalIFuncCallOrRef; 650 GlobalValueSummary::GVFlags Flags( 651 F.getLinkage(), F.getVisibility(), NotEligibleForImport, 652 /* Live = */ false, F.isDSOLocal(), F.canBeOmittedFromSymbolTable(), 653 GlobalValueSummary::ImportKind::Definition); 654 FunctionSummary::FFlags FunFlags{ 655 F.doesNotAccessMemory(), F.onlyReadsMemory() && !F.doesNotAccessMemory(), 656 F.hasFnAttribute(Attribute::NoRecurse), F.returnDoesNotAlias(), 657 // FIXME: refactor this to use the same code that inliner is using. 658 // Don't try to import functions with noinline attribute. 659 F.getAttributes().hasFnAttr(Attribute::NoInline), 660 F.hasFnAttribute(Attribute::AlwaysInline), 661 F.hasFnAttribute(Attribute::NoUnwind), MayThrow, HasUnknownCall, 662 mustBeUnreachableFunction(F)}; 663 std::vector<FunctionSummary::ParamAccess> ParamAccesses; 664 if (auto *SSI = GetSSICallback(F)) 665 ParamAccesses = SSI->getParamAccesses(Index); 666 auto FuncSummary = std::make_unique<FunctionSummary>( 667 Flags, NumInsts, FunFlags, /*EntryCount=*/0, std::move(Refs), 668 CallGraphEdges.takeVector(), TypeTests.takeVector(), 669 TypeTestAssumeVCalls.takeVector(), TypeCheckedLoadVCalls.takeVector(), 670 TypeTestAssumeConstVCalls.takeVector(), 671 TypeCheckedLoadConstVCalls.takeVector(), std::move(ParamAccesses), 672 std::move(Callsites), std::move(Allocs)); 673 if (NonRenamableLocal) 674 CantBePromoted.insert(F.getGUID()); 675 Index.addGlobalValueSummary(F, std::move(FuncSummary)); 676 } 677 678 /// Find function pointers referenced within the given vtable initializer 679 /// (or subset of an initializer) \p I. The starting offset of \p I within 680 /// the vtable initializer is \p StartingOffset. Any discovered function 681 /// pointers are added to \p VTableFuncs along with their cumulative offset 682 /// within the initializer. 683 static void findFuncPointers(const Constant *I, uint64_t StartingOffset, 684 const Module &M, ModuleSummaryIndex &Index, 685 VTableFuncList &VTableFuncs, 686 const GlobalVariable &OrigGV) { 687 // First check if this is a function pointer. 688 if (I->getType()->isPointerTy()) { 689 auto C = I->stripPointerCasts(); 690 auto A = dyn_cast<GlobalAlias>(C); 691 if (isa<Function>(C) || (A && isa<Function>(A->getAliasee()))) { 692 auto GV = dyn_cast<GlobalValue>(C); 693 assert(GV); 694 // We can disregard __cxa_pure_virtual as a possible call target, as 695 // calls to pure virtuals are UB. 696 if (GV && GV->getName() != "__cxa_pure_virtual") 697 VTableFuncs.push_back({Index.getOrInsertValueInfo(GV), StartingOffset}); 698 return; 699 } 700 } 701 702 // Walk through the elements in the constant struct or array and recursively 703 // look for virtual function pointers. 704 const DataLayout &DL = M.getDataLayout(); 705 if (auto *C = dyn_cast<ConstantStruct>(I)) { 706 StructType *STy = dyn_cast<StructType>(C->getType()); 707 assert(STy); 708 const StructLayout *SL = DL.getStructLayout(C->getType()); 709 710 for (auto EI : llvm::enumerate(STy->elements())) { 711 auto Offset = SL->getElementOffset(EI.index()); 712 unsigned Op = SL->getElementContainingOffset(Offset); 713 findFuncPointers(cast<Constant>(I->getOperand(Op)), 714 StartingOffset + Offset, M, Index, VTableFuncs, OrigGV); 715 } 716 } else if (auto *C = dyn_cast<ConstantArray>(I)) { 717 ArrayType *ATy = C->getType(); 718 Type *EltTy = ATy->getElementType(); 719 uint64_t EltSize = DL.getTypeAllocSize(EltTy); 720 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) { 721 findFuncPointers(cast<Constant>(I->getOperand(i)), 722 StartingOffset + i * EltSize, M, Index, VTableFuncs, 723 OrigGV); 724 } 725 } else if (const auto *CE = dyn_cast<ConstantExpr>(I)) { 726 // For relative vtables, the next sub-component should be a trunc. 727 if (CE->getOpcode() != Instruction::Trunc || 728 !(CE = dyn_cast<ConstantExpr>(CE->getOperand(0)))) 729 return; 730 731 // If this constant can be reduced to the offset between a function and a 732 // global, then we know this is a valid virtual function if the RHS is the 733 // original vtable we're scanning through. 734 if (CE->getOpcode() == Instruction::Sub) { 735 GlobalValue *LHS, *RHS; 736 APSInt LHSOffset, RHSOffset; 737 if (IsConstantOffsetFromGlobal(CE->getOperand(0), LHS, LHSOffset, DL) && 738 IsConstantOffsetFromGlobal(CE->getOperand(1), RHS, RHSOffset, DL) && 739 RHS == &OrigGV && 740 741 // For relative vtables, this component should point to the callable 742 // function without any offsets. 743 LHSOffset == 0 && 744 745 // Also, the RHS should always point to somewhere within the vtable. 746 RHSOffset <= 747 static_cast<uint64_t>(DL.getTypeAllocSize(OrigGV.getInitializer()->getType()))) { 748 findFuncPointers(LHS, StartingOffset, M, Index, VTableFuncs, OrigGV); 749 } 750 } 751 } 752 } 753 754 // Identify the function pointers referenced by vtable definition \p V. 755 static void computeVTableFuncs(ModuleSummaryIndex &Index, 756 const GlobalVariable &V, const Module &M, 757 VTableFuncList &VTableFuncs) { 758 if (!V.isConstant()) 759 return; 760 761 findFuncPointers(V.getInitializer(), /*StartingOffset=*/0, M, Index, 762 VTableFuncs, V); 763 764 #ifndef NDEBUG 765 // Validate that the VTableFuncs list is ordered by offset. 766 uint64_t PrevOffset = 0; 767 for (auto &P : VTableFuncs) { 768 // The findVFuncPointers traversal should have encountered the 769 // functions in offset order. We need to use ">=" since PrevOffset 770 // starts at 0. 771 assert(P.VTableOffset >= PrevOffset); 772 PrevOffset = P.VTableOffset; 773 } 774 #endif 775 } 776 777 /// Record vtable definition \p V for each type metadata it references. 778 static void 779 recordTypeIdCompatibleVtableReferences(ModuleSummaryIndex &Index, 780 const GlobalVariable &V, 781 SmallVectorImpl<MDNode *> &Types) { 782 for (MDNode *Type : Types) { 783 auto TypeID = Type->getOperand(1).get(); 784 785 uint64_t Offset = 786 cast<ConstantInt>( 787 cast<ConstantAsMetadata>(Type->getOperand(0))->getValue()) 788 ->getZExtValue(); 789 790 if (auto *TypeId = dyn_cast<MDString>(TypeID)) 791 Index.getOrInsertTypeIdCompatibleVtableSummary(TypeId->getString()) 792 .push_back({Offset, Index.getOrInsertValueInfo(&V)}); 793 } 794 } 795 796 static void computeVariableSummary(ModuleSummaryIndex &Index, 797 const GlobalVariable &V, 798 DenseSet<GlobalValue::GUID> &CantBePromoted, 799 const Module &M, 800 SmallVectorImpl<MDNode *> &Types) { 801 SetVector<ValueInfo, std::vector<ValueInfo>> RefEdges; 802 SmallPtrSet<const User *, 8> Visited; 803 bool RefLocalIFunc = false; 804 bool HasBlockAddress = 805 findRefEdges(Index, &V, RefEdges, Visited, RefLocalIFunc); 806 const bool NotEligibleForImport = (HasBlockAddress || RefLocalIFunc); 807 bool NonRenamableLocal = isNonRenamableLocal(V); 808 GlobalValueSummary::GVFlags Flags( 809 V.getLinkage(), V.getVisibility(), NonRenamableLocal, 810 /* Live = */ false, V.isDSOLocal(), V.canBeOmittedFromSymbolTable(), 811 GlobalValueSummary::Definition); 812 813 VTableFuncList VTableFuncs; 814 // If splitting is not enabled, then we compute the summary information 815 // necessary for index-based whole program devirtualization. 816 if (!Index.enableSplitLTOUnit()) { 817 Types.clear(); 818 V.getMetadata(LLVMContext::MD_type, Types); 819 if (!Types.empty()) { 820 // Identify the function pointers referenced by this vtable definition. 821 computeVTableFuncs(Index, V, M, VTableFuncs); 822 823 // Record this vtable definition for each type metadata it references. 824 recordTypeIdCompatibleVtableReferences(Index, V, Types); 825 } 826 } 827 828 // Don't mark variables we won't be able to internalize as read/write-only. 829 bool CanBeInternalized = 830 !V.hasComdat() && !V.hasAppendingLinkage() && !V.isInterposable() && 831 !V.hasAvailableExternallyLinkage() && !V.hasDLLExportStorageClass(); 832 bool Constant = V.isConstant(); 833 GlobalVarSummary::GVarFlags VarFlags(CanBeInternalized, 834 Constant ? false : CanBeInternalized, 835 Constant, V.getVCallVisibility()); 836 auto GVarSummary = std::make_unique<GlobalVarSummary>(Flags, VarFlags, 837 RefEdges.takeVector()); 838 if (NonRenamableLocal) 839 CantBePromoted.insert(V.getGUID()); 840 if (NotEligibleForImport) 841 GVarSummary->setNotEligibleToImport(); 842 if (!VTableFuncs.empty()) 843 GVarSummary->setVTableFuncs(VTableFuncs); 844 Index.addGlobalValueSummary(V, std::move(GVarSummary)); 845 } 846 847 static void computeAliasSummary(ModuleSummaryIndex &Index, const GlobalAlias &A, 848 DenseSet<GlobalValue::GUID> &CantBePromoted) { 849 // Skip summary for indirect function aliases as summary for aliasee will not 850 // be emitted. 851 const GlobalObject *Aliasee = A.getAliaseeObject(); 852 if (isa<GlobalIFunc>(Aliasee)) 853 return; 854 bool NonRenamableLocal = isNonRenamableLocal(A); 855 GlobalValueSummary::GVFlags Flags( 856 A.getLinkage(), A.getVisibility(), NonRenamableLocal, 857 /* Live = */ false, A.isDSOLocal(), A.canBeOmittedFromSymbolTable(), 858 GlobalValueSummary::Definition); 859 auto AS = std::make_unique<AliasSummary>(Flags); 860 auto AliaseeVI = Index.getValueInfo(Aliasee->getGUID()); 861 assert(AliaseeVI && "Alias expects aliasee summary to be available"); 862 assert(AliaseeVI.getSummaryList().size() == 1 && 863 "Expected a single entry per aliasee in per-module index"); 864 AS->setAliasee(AliaseeVI, AliaseeVI.getSummaryList()[0].get()); 865 if (NonRenamableLocal) 866 CantBePromoted.insert(A.getGUID()); 867 Index.addGlobalValueSummary(A, std::move(AS)); 868 } 869 870 // Set LiveRoot flag on entries matching the given value name. 871 static void setLiveRoot(ModuleSummaryIndex &Index, StringRef Name) { 872 if (ValueInfo VI = Index.getValueInfo(GlobalValue::getGUID(Name))) 873 for (const auto &Summary : VI.getSummaryList()) 874 Summary->setLive(true); 875 } 876 877 ModuleSummaryIndex llvm::buildModuleSummaryIndex( 878 const Module &M, 879 std::function<BlockFrequencyInfo *(const Function &F)> GetBFICallback, 880 ProfileSummaryInfo *PSI, 881 std::function<const StackSafetyInfo *(const Function &F)> GetSSICallback) { 882 assert(PSI); 883 bool EnableSplitLTOUnit = false; 884 bool UnifiedLTO = false; 885 if (auto *MD = mdconst::extract_or_null<ConstantInt>( 886 M.getModuleFlag("EnableSplitLTOUnit"))) 887 EnableSplitLTOUnit = MD->getZExtValue(); 888 if (auto *MD = 889 mdconst::extract_or_null<ConstantInt>(M.getModuleFlag("UnifiedLTO"))) 890 UnifiedLTO = MD->getZExtValue(); 891 ModuleSummaryIndex Index(/*HaveGVs=*/true, EnableSplitLTOUnit, UnifiedLTO); 892 893 // Identify the local values in the llvm.used and llvm.compiler.used sets, 894 // which should not be exported as they would then require renaming and 895 // promotion, but we may have opaque uses e.g. in inline asm. We collect them 896 // here because we use this information to mark functions containing inline 897 // assembly calls as not importable. 898 SmallPtrSet<GlobalValue *, 4> LocalsUsed; 899 SmallVector<GlobalValue *, 4> Used; 900 // First collect those in the llvm.used set. 901 collectUsedGlobalVariables(M, Used, /*CompilerUsed=*/false); 902 // Next collect those in the llvm.compiler.used set. 903 collectUsedGlobalVariables(M, Used, /*CompilerUsed=*/true); 904 DenseSet<GlobalValue::GUID> CantBePromoted; 905 for (auto *V : Used) { 906 if (V->hasLocalLinkage()) { 907 LocalsUsed.insert(V); 908 CantBePromoted.insert(V->getGUID()); 909 } 910 } 911 912 bool HasLocalInlineAsmSymbol = false; 913 if (!M.getModuleInlineAsm().empty()) { 914 // Collect the local values defined by module level asm, and set up 915 // summaries for these symbols so that they can be marked as NoRename, 916 // to prevent export of any use of them in regular IR that would require 917 // renaming within the module level asm. Note we don't need to create a 918 // summary for weak or global defs, as they don't need to be flagged as 919 // NoRename, and defs in module level asm can't be imported anyway. 920 // Also, any values used but not defined within module level asm should 921 // be listed on the llvm.used or llvm.compiler.used global and marked as 922 // referenced from there. 923 ModuleSymbolTable::CollectAsmSymbols( 924 M, [&](StringRef Name, object::BasicSymbolRef::Flags Flags) { 925 // Symbols not marked as Weak or Global are local definitions. 926 if (Flags & (object::BasicSymbolRef::SF_Weak | 927 object::BasicSymbolRef::SF_Global)) 928 return; 929 HasLocalInlineAsmSymbol = true; 930 GlobalValue *GV = M.getNamedValue(Name); 931 if (!GV) 932 return; 933 assert(GV->isDeclaration() && "Def in module asm already has definition"); 934 GlobalValueSummary::GVFlags GVFlags( 935 GlobalValue::InternalLinkage, GlobalValue::DefaultVisibility, 936 /* NotEligibleToImport = */ true, 937 /* Live = */ true, 938 /* Local */ GV->isDSOLocal(), GV->canBeOmittedFromSymbolTable(), 939 GlobalValueSummary::Definition); 940 CantBePromoted.insert(GV->getGUID()); 941 // Create the appropriate summary type. 942 if (Function *F = dyn_cast<Function>(GV)) { 943 std::unique_ptr<FunctionSummary> Summary = 944 std::make_unique<FunctionSummary>( 945 GVFlags, /*InstCount=*/0, 946 FunctionSummary::FFlags{ 947 F->hasFnAttribute(Attribute::ReadNone), 948 F->hasFnAttribute(Attribute::ReadOnly), 949 F->hasFnAttribute(Attribute::NoRecurse), 950 F->returnDoesNotAlias(), 951 /* NoInline = */ false, 952 F->hasFnAttribute(Attribute::AlwaysInline), 953 F->hasFnAttribute(Attribute::NoUnwind), 954 /* MayThrow */ true, 955 /* HasUnknownCall */ true, 956 /* MustBeUnreachable */ false}, 957 /*EntryCount=*/0, ArrayRef<ValueInfo>{}, 958 ArrayRef<FunctionSummary::EdgeTy>{}, 959 ArrayRef<GlobalValue::GUID>{}, 960 ArrayRef<FunctionSummary::VFuncId>{}, 961 ArrayRef<FunctionSummary::VFuncId>{}, 962 ArrayRef<FunctionSummary::ConstVCall>{}, 963 ArrayRef<FunctionSummary::ConstVCall>{}, 964 ArrayRef<FunctionSummary::ParamAccess>{}, 965 ArrayRef<CallsiteInfo>{}, ArrayRef<AllocInfo>{}); 966 Index.addGlobalValueSummary(*GV, std::move(Summary)); 967 } else { 968 std::unique_ptr<GlobalVarSummary> Summary = 969 std::make_unique<GlobalVarSummary>( 970 GVFlags, 971 GlobalVarSummary::GVarFlags( 972 false, false, cast<GlobalVariable>(GV)->isConstant(), 973 GlobalObject::VCallVisibilityPublic), 974 ArrayRef<ValueInfo>{}); 975 Index.addGlobalValueSummary(*GV, std::move(Summary)); 976 } 977 }); 978 } 979 980 bool IsThinLTO = true; 981 if (auto *MD = 982 mdconst::extract_or_null<ConstantInt>(M.getModuleFlag("ThinLTO"))) 983 IsThinLTO = MD->getZExtValue(); 984 985 // Compute summaries for all functions defined in module, and save in the 986 // index. 987 for (const auto &F : M) { 988 if (F.isDeclaration()) 989 continue; 990 991 DominatorTree DT(const_cast<Function &>(F)); 992 BlockFrequencyInfo *BFI = nullptr; 993 std::unique_ptr<BlockFrequencyInfo> BFIPtr; 994 if (GetBFICallback) 995 BFI = GetBFICallback(F); 996 else if (F.hasProfileData()) { 997 LoopInfo LI{DT}; 998 BranchProbabilityInfo BPI{F, LI}; 999 BFIPtr = std::make_unique<BlockFrequencyInfo>(F, BPI, LI); 1000 BFI = BFIPtr.get(); 1001 } 1002 1003 computeFunctionSummary(Index, M, F, BFI, PSI, DT, 1004 !LocalsUsed.empty() || HasLocalInlineAsmSymbol, 1005 CantBePromoted, IsThinLTO, GetSSICallback); 1006 } 1007 1008 // Compute summaries for all variables defined in module, and save in the 1009 // index. 1010 SmallVector<MDNode *, 2> Types; 1011 for (const GlobalVariable &G : M.globals()) { 1012 if (G.isDeclaration()) 1013 continue; 1014 computeVariableSummary(Index, G, CantBePromoted, M, Types); 1015 } 1016 1017 // Compute summaries for all aliases defined in module, and save in the 1018 // index. 1019 for (const GlobalAlias &A : M.aliases()) 1020 computeAliasSummary(Index, A, CantBePromoted); 1021 1022 // Iterate through ifuncs, set their resolvers all alive. 1023 for (const GlobalIFunc &I : M.ifuncs()) { 1024 I.applyAlongResolverPath([&Index](const GlobalValue &GV) { 1025 Index.getGlobalValueSummary(GV)->setLive(true); 1026 }); 1027 } 1028 1029 for (auto *V : LocalsUsed) { 1030 auto *Summary = Index.getGlobalValueSummary(*V); 1031 assert(Summary && "Missing summary for global value"); 1032 Summary->setNotEligibleToImport(); 1033 } 1034 1035 // The linker doesn't know about these LLVM produced values, so we need 1036 // to flag them as live in the index to ensure index-based dead value 1037 // analysis treats them as live roots of the analysis. 1038 setLiveRoot(Index, "llvm.used"); 1039 setLiveRoot(Index, "llvm.compiler.used"); 1040 setLiveRoot(Index, "llvm.global_ctors"); 1041 setLiveRoot(Index, "llvm.global_dtors"); 1042 setLiveRoot(Index, "llvm.global.annotations"); 1043 1044 for (auto &GlobalList : Index) { 1045 // Ignore entries for references that are undefined in the current module. 1046 if (GlobalList.second.SummaryList.empty()) 1047 continue; 1048 1049 assert(GlobalList.second.SummaryList.size() == 1 && 1050 "Expected module's index to have one summary per GUID"); 1051 auto &Summary = GlobalList.second.SummaryList[0]; 1052 if (!IsThinLTO) { 1053 Summary->setNotEligibleToImport(); 1054 continue; 1055 } 1056 1057 bool AllRefsCanBeExternallyReferenced = 1058 llvm::all_of(Summary->refs(), [&](const ValueInfo &VI) { 1059 return !CantBePromoted.count(VI.getGUID()); 1060 }); 1061 if (!AllRefsCanBeExternallyReferenced) { 1062 Summary->setNotEligibleToImport(); 1063 continue; 1064 } 1065 1066 if (auto *FuncSummary = dyn_cast<FunctionSummary>(Summary.get())) { 1067 bool AllCallsCanBeExternallyReferenced = llvm::all_of( 1068 FuncSummary->calls(), [&](const FunctionSummary::EdgeTy &Edge) { 1069 return !CantBePromoted.count(Edge.first.getGUID()); 1070 }); 1071 if (!AllCallsCanBeExternallyReferenced) 1072 Summary->setNotEligibleToImport(); 1073 } 1074 } 1075 1076 if (!ModuleSummaryDotFile.empty()) { 1077 std::error_code EC; 1078 raw_fd_ostream OSDot(ModuleSummaryDotFile, EC, sys::fs::OpenFlags::OF_Text); 1079 if (EC) 1080 report_fatal_error(Twine("Failed to open dot file ") + 1081 ModuleSummaryDotFile + ": " + EC.message() + "\n"); 1082 Index.exportToDot(OSDot, {}); 1083 } 1084 1085 return Index; 1086 } 1087 1088 AnalysisKey ModuleSummaryIndexAnalysis::Key; 1089 1090 ModuleSummaryIndex 1091 ModuleSummaryIndexAnalysis::run(Module &M, ModuleAnalysisManager &AM) { 1092 ProfileSummaryInfo &PSI = AM.getResult<ProfileSummaryAnalysis>(M); 1093 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 1094 bool NeedSSI = needsParamAccessSummary(M); 1095 return buildModuleSummaryIndex( 1096 M, 1097 [&FAM](const Function &F) { 1098 return &FAM.getResult<BlockFrequencyAnalysis>( 1099 *const_cast<Function *>(&F)); 1100 }, 1101 &PSI, 1102 [&FAM, NeedSSI](const Function &F) -> const StackSafetyInfo * { 1103 return NeedSSI ? &FAM.getResult<StackSafetyAnalysis>( 1104 const_cast<Function &>(F)) 1105 : nullptr; 1106 }); 1107 } 1108 1109 char ModuleSummaryIndexWrapperPass::ID = 0; 1110 1111 INITIALIZE_PASS_BEGIN(ModuleSummaryIndexWrapperPass, "module-summary-analysis", 1112 "Module Summary Analysis", false, true) 1113 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) 1114 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass) 1115 INITIALIZE_PASS_DEPENDENCY(StackSafetyInfoWrapperPass) 1116 INITIALIZE_PASS_END(ModuleSummaryIndexWrapperPass, "module-summary-analysis", 1117 "Module Summary Analysis", false, true) 1118 1119 ModulePass *llvm::createModuleSummaryIndexWrapperPass() { 1120 return new ModuleSummaryIndexWrapperPass(); 1121 } 1122 1123 ModuleSummaryIndexWrapperPass::ModuleSummaryIndexWrapperPass() 1124 : ModulePass(ID) { 1125 initializeModuleSummaryIndexWrapperPassPass(*PassRegistry::getPassRegistry()); 1126 } 1127 1128 bool ModuleSummaryIndexWrapperPass::runOnModule(Module &M) { 1129 auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 1130 bool NeedSSI = needsParamAccessSummary(M); 1131 Index.emplace(buildModuleSummaryIndex( 1132 M, 1133 [this](const Function &F) { 1134 return &(this->getAnalysis<BlockFrequencyInfoWrapperPass>( 1135 *const_cast<Function *>(&F)) 1136 .getBFI()); 1137 }, 1138 PSI, 1139 [&](const Function &F) -> const StackSafetyInfo * { 1140 return NeedSSI ? &getAnalysis<StackSafetyInfoWrapperPass>( 1141 const_cast<Function &>(F)) 1142 .getResult() 1143 : nullptr; 1144 })); 1145 return false; 1146 } 1147 1148 bool ModuleSummaryIndexWrapperPass::doFinalization(Module &M) { 1149 Index.reset(); 1150 return false; 1151 } 1152 1153 void ModuleSummaryIndexWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 1154 AU.setPreservesAll(); 1155 AU.addRequired<BlockFrequencyInfoWrapperPass>(); 1156 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 1157 AU.addRequired<StackSafetyInfoWrapperPass>(); 1158 } 1159 1160 char ImmutableModuleSummaryIndexWrapperPass::ID = 0; 1161 1162 ImmutableModuleSummaryIndexWrapperPass::ImmutableModuleSummaryIndexWrapperPass( 1163 const ModuleSummaryIndex *Index) 1164 : ImmutablePass(ID), Index(Index) { 1165 initializeImmutableModuleSummaryIndexWrapperPassPass( 1166 *PassRegistry::getPassRegistry()); 1167 } 1168 1169 void ImmutableModuleSummaryIndexWrapperPass::getAnalysisUsage( 1170 AnalysisUsage &AU) const { 1171 AU.setPreservesAll(); 1172 } 1173 1174 ImmutablePass *llvm::createImmutableModuleSummaryIndexWrapperPass( 1175 const ModuleSummaryIndex *Index) { 1176 return new ImmutableModuleSummaryIndexWrapperPass(Index); 1177 } 1178 1179 INITIALIZE_PASS(ImmutableModuleSummaryIndexWrapperPass, "module-summary-info", 1180 "Module summary info", false, true) 1181 1182 bool llvm::mayHaveMemprofSummary(const CallBase *CB) { 1183 if (!CB) 1184 return false; 1185 if (CB->isDebugOrPseudoInst()) 1186 return false; 1187 auto *CI = dyn_cast<CallInst>(CB); 1188 auto *CalledValue = CB->getCalledOperand(); 1189 auto *CalledFunction = CB->getCalledFunction(); 1190 if (CalledValue && !CalledFunction) { 1191 CalledValue = CalledValue->stripPointerCasts(); 1192 // Stripping pointer casts can reveal a called function. 1193 CalledFunction = dyn_cast<Function>(CalledValue); 1194 } 1195 // Check if this is an alias to a function. If so, get the 1196 // called aliasee for the checks below. 1197 if (auto *GA = dyn_cast<GlobalAlias>(CalledValue)) { 1198 assert(!CalledFunction && 1199 "Expected null called function in callsite for alias"); 1200 CalledFunction = dyn_cast<Function>(GA->getAliaseeObject()); 1201 } 1202 // Check if this is a direct call to a known function or a known 1203 // intrinsic, or an indirect call with profile data. 1204 if (CalledFunction) { 1205 if (CI && CalledFunction->isIntrinsic()) 1206 return false; 1207 } else { 1208 // TODO: For now skip indirect calls. See comments in 1209 // computeFunctionSummary for what is needed to handle this. 1210 return false; 1211 } 1212 return true; 1213 } 1214