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