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