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/TypeMetadataUtils.h" 29 #include "llvm/IR/Attributes.h" 30 #include "llvm/IR/BasicBlock.h" 31 #include "llvm/IR/CallSite.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/Object/ModuleSymbolTable.h" 48 #include "llvm/Object/SymbolicFile.h" 49 #include "llvm/Pass.h" 50 #include "llvm/Support/Casting.h" 51 #include "llvm/Support/CommandLine.h" 52 #include <algorithm> 53 #include <cassert> 54 #include <cstdint> 55 #include <vector> 56 57 using namespace llvm; 58 59 #define DEBUG_TYPE "module-summary-analysis" 60 61 // Option to force edges cold which will block importing when the 62 // -import-cold-multiplier is set to 0. Useful for debugging. 63 FunctionSummary::ForceSummaryHotnessType ForceSummaryEdgesCold = 64 FunctionSummary::FSHT_None; 65 cl::opt<FunctionSummary::ForceSummaryHotnessType, true> FSEC( 66 "force-summary-edges-cold", cl::Hidden, cl::location(ForceSummaryEdgesCold), 67 cl::desc("Force all edges in the function summary to cold"), 68 cl::values(clEnumValN(FunctionSummary::FSHT_None, "none", "None."), 69 clEnumValN(FunctionSummary::FSHT_AllNonCritical, 70 "all-non-critical", "All non-critical edges."), 71 clEnumValN(FunctionSummary::FSHT_All, "all", "All edges."))); 72 73 cl::opt<std::string> ModuleSummaryDotFile( 74 "module-summary-dot-file", cl::init(""), cl::Hidden, 75 cl::value_desc("filename"), 76 cl::desc("File to emit dot graph of new summary into.")); 77 78 // Walk through the operands of a given User via worklist iteration and populate 79 // the set of GlobalValue references encountered. Invoked either on an 80 // Instruction or a GlobalVariable (which walks its initializer). 81 // Return true if any of the operands contains blockaddress. This is important 82 // to know when computing summary for global var, because if global variable 83 // references basic block address we can't import it separately from function 84 // containing that basic block. For simplicity we currently don't import such 85 // global vars at all. When importing function we aren't interested if any 86 // instruction in it takes an address of any basic block, because instruction 87 // can only take an address of basic block located in the same function. 88 static bool findRefEdges(ModuleSummaryIndex &Index, const User *CurUser, 89 SetVector<ValueInfo> &RefEdges, 90 SmallPtrSet<const User *, 8> &Visited) { 91 bool HasBlockAddress = false; 92 SmallVector<const User *, 32> Worklist; 93 Worklist.push_back(CurUser); 94 95 while (!Worklist.empty()) { 96 const User *U = Worklist.pop_back_val(); 97 98 if (!Visited.insert(U).second) 99 continue; 100 101 ImmutableCallSite CS(U); 102 103 for (const auto &OI : U->operands()) { 104 const User *Operand = dyn_cast<User>(OI); 105 if (!Operand) 106 continue; 107 if (isa<BlockAddress>(Operand)) { 108 HasBlockAddress = true; 109 continue; 110 } 111 if (auto *GV = dyn_cast<GlobalValue>(Operand)) { 112 // We have a reference to a global value. This should be added to 113 // the reference set unless it is a callee. Callees are handled 114 // specially by WriteFunction and are added to a separate list. 115 if (!(CS && CS.isCallee(&OI))) 116 RefEdges.insert(Index.getOrInsertValueInfo(GV)); 117 continue; 118 } 119 Worklist.push_back(Operand); 120 } 121 } 122 return HasBlockAddress; 123 } 124 125 static CalleeInfo::HotnessType getHotness(uint64_t ProfileCount, 126 ProfileSummaryInfo *PSI) { 127 if (!PSI) 128 return CalleeInfo::HotnessType::Unknown; 129 if (PSI->isHotCount(ProfileCount)) 130 return CalleeInfo::HotnessType::Hot; 131 if (PSI->isColdCount(ProfileCount)) 132 return CalleeInfo::HotnessType::Cold; 133 return CalleeInfo::HotnessType::None; 134 } 135 136 static bool isNonRenamableLocal(const GlobalValue &GV) { 137 return GV.hasSection() && GV.hasLocalLinkage(); 138 } 139 140 /// Determine whether this call has all constant integer arguments (excluding 141 /// "this") and summarize it to VCalls or ConstVCalls as appropriate. 142 static void addVCallToSet(DevirtCallSite Call, GlobalValue::GUID Guid, 143 SetVector<FunctionSummary::VFuncId> &VCalls, 144 SetVector<FunctionSummary::ConstVCall> &ConstVCalls) { 145 std::vector<uint64_t> Args; 146 // Start from the second argument to skip the "this" pointer. 147 for (auto &Arg : make_range(Call.CS.arg_begin() + 1, Call.CS.arg_end())) { 148 auto *CI = dyn_cast<ConstantInt>(Arg); 149 if (!CI || CI->getBitWidth() > 64) { 150 VCalls.insert({Guid, Call.Offset}); 151 return; 152 } 153 Args.push_back(CI->getZExtValue()); 154 } 155 ConstVCalls.insert({{Guid, Call.Offset}, std::move(Args)}); 156 } 157 158 /// If this intrinsic call requires that we add information to the function 159 /// summary, do so via the non-constant reference arguments. 160 static void addIntrinsicToSummary( 161 const CallInst *CI, SetVector<GlobalValue::GUID> &TypeTests, 162 SetVector<FunctionSummary::VFuncId> &TypeTestAssumeVCalls, 163 SetVector<FunctionSummary::VFuncId> &TypeCheckedLoadVCalls, 164 SetVector<FunctionSummary::ConstVCall> &TypeTestAssumeConstVCalls, 165 SetVector<FunctionSummary::ConstVCall> &TypeCheckedLoadConstVCalls, 166 DominatorTree &DT) { 167 switch (CI->getCalledFunction()->getIntrinsicID()) { 168 case Intrinsic::type_test: { 169 auto *TypeMDVal = cast<MetadataAsValue>(CI->getArgOperand(1)); 170 auto *TypeId = dyn_cast<MDString>(TypeMDVal->getMetadata()); 171 if (!TypeId) 172 break; 173 GlobalValue::GUID Guid = GlobalValue::getGUID(TypeId->getString()); 174 175 // Produce a summary from type.test intrinsics. We only summarize type.test 176 // intrinsics that are used other than by an llvm.assume intrinsic. 177 // Intrinsics that are assumed are relevant only to the devirtualization 178 // pass, not the type test lowering pass. 179 bool HasNonAssumeUses = llvm::any_of(CI->uses(), [](const Use &CIU) { 180 auto *AssumeCI = dyn_cast<CallInst>(CIU.getUser()); 181 if (!AssumeCI) 182 return true; 183 Function *F = AssumeCI->getCalledFunction(); 184 return !F || F->getIntrinsicID() != Intrinsic::assume; 185 }); 186 if (HasNonAssumeUses) 187 TypeTests.insert(Guid); 188 189 SmallVector<DevirtCallSite, 4> DevirtCalls; 190 SmallVector<CallInst *, 4> Assumes; 191 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT); 192 for (auto &Call : DevirtCalls) 193 addVCallToSet(Call, Guid, TypeTestAssumeVCalls, 194 TypeTestAssumeConstVCalls); 195 196 break; 197 } 198 199 case Intrinsic::type_checked_load: { 200 auto *TypeMDVal = cast<MetadataAsValue>(CI->getArgOperand(2)); 201 auto *TypeId = dyn_cast<MDString>(TypeMDVal->getMetadata()); 202 if (!TypeId) 203 break; 204 GlobalValue::GUID Guid = GlobalValue::getGUID(TypeId->getString()); 205 206 SmallVector<DevirtCallSite, 4> DevirtCalls; 207 SmallVector<Instruction *, 4> LoadedPtrs; 208 SmallVector<Instruction *, 4> Preds; 209 bool HasNonCallUses = false; 210 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds, 211 HasNonCallUses, CI, DT); 212 // Any non-call uses of the result of llvm.type.checked.load will 213 // prevent us from optimizing away the llvm.type.test. 214 if (HasNonCallUses) 215 TypeTests.insert(Guid); 216 for (auto &Call : DevirtCalls) 217 addVCallToSet(Call, Guid, TypeCheckedLoadVCalls, 218 TypeCheckedLoadConstVCalls); 219 220 break; 221 } 222 default: 223 break; 224 } 225 } 226 227 static bool isNonVolatileLoad(const Instruction *I) { 228 if (const auto *LI = dyn_cast<LoadInst>(I)) 229 return !LI->isVolatile(); 230 231 return false; 232 } 233 234 static void computeFunctionSummary(ModuleSummaryIndex &Index, const Module &M, 235 const Function &F, BlockFrequencyInfo *BFI, 236 ProfileSummaryInfo *PSI, DominatorTree &DT, 237 bool HasLocalsInUsedOrAsm, 238 DenseSet<GlobalValue::GUID> &CantBePromoted, 239 bool IsThinLTO) { 240 // Summary not currently supported for anonymous functions, they should 241 // have been named. 242 assert(F.hasName()); 243 244 unsigned NumInsts = 0; 245 // Map from callee ValueId to profile count. Used to accumulate profile 246 // counts for all static calls to a given callee. 247 MapVector<ValueInfo, CalleeInfo> CallGraphEdges; 248 SetVector<ValueInfo> RefEdges; 249 SetVector<GlobalValue::GUID> TypeTests; 250 SetVector<FunctionSummary::VFuncId> TypeTestAssumeVCalls, 251 TypeCheckedLoadVCalls; 252 SetVector<FunctionSummary::ConstVCall> TypeTestAssumeConstVCalls, 253 TypeCheckedLoadConstVCalls; 254 ICallPromotionAnalysis ICallAnalysis; 255 SmallPtrSet<const User *, 8> Visited; 256 257 // Add personality function, prefix data and prologue data to function's ref 258 // list. 259 findRefEdges(Index, &F, RefEdges, Visited); 260 std::vector<const Instruction *> NonVolatileLoads; 261 262 bool HasInlineAsmMaybeReferencingInternal = false; 263 for (const BasicBlock &BB : F) 264 for (const Instruction &I : BB) { 265 if (isa<DbgInfoIntrinsic>(I)) 266 continue; 267 ++NumInsts; 268 if (isNonVolatileLoad(&I)) { 269 // Postpone processing of non-volatile load instructions 270 // See comments below 271 Visited.insert(&I); 272 NonVolatileLoads.push_back(&I); 273 continue; 274 } 275 findRefEdges(Index, &I, RefEdges, Visited); 276 auto CS = ImmutableCallSite(&I); 277 if (!CS) 278 continue; 279 280 const auto *CI = dyn_cast<CallInst>(&I); 281 // Since we don't know exactly which local values are referenced in inline 282 // assembly, conservatively mark the function as possibly referencing 283 // a local value from inline assembly to ensure we don't export a 284 // reference (which would require renaming and promotion of the 285 // referenced value). 286 if (HasLocalsInUsedOrAsm && CI && CI->isInlineAsm()) 287 HasInlineAsmMaybeReferencingInternal = true; 288 289 auto *CalledValue = CS.getCalledValue(); 290 auto *CalledFunction = CS.getCalledFunction(); 291 if (CalledValue && !CalledFunction) { 292 CalledValue = CalledValue->stripPointerCastsNoFollowAliases(); 293 // Stripping pointer casts can reveal a called function. 294 CalledFunction = dyn_cast<Function>(CalledValue); 295 } 296 // Check if this is an alias to a function. If so, get the 297 // called aliasee for the checks below. 298 if (auto *GA = dyn_cast<GlobalAlias>(CalledValue)) { 299 assert(!CalledFunction && "Expected null called function in callsite for alias"); 300 CalledFunction = dyn_cast<Function>(GA->getBaseObject()); 301 } 302 // Check if this is a direct call to a known function or a known 303 // intrinsic, or an indirect call with profile data. 304 if (CalledFunction) { 305 if (CI && CalledFunction->isIntrinsic()) { 306 addIntrinsicToSummary( 307 CI, TypeTests, TypeTestAssumeVCalls, TypeCheckedLoadVCalls, 308 TypeTestAssumeConstVCalls, TypeCheckedLoadConstVCalls, DT); 309 continue; 310 } 311 // We should have named any anonymous globals 312 assert(CalledFunction->hasName()); 313 auto ScaledCount = PSI->getProfileCount(&I, BFI); 314 auto Hotness = ScaledCount ? getHotness(ScaledCount.getValue(), PSI) 315 : CalleeInfo::HotnessType::Unknown; 316 if (ForceSummaryEdgesCold != FunctionSummary::FSHT_None) 317 Hotness = CalleeInfo::HotnessType::Cold; 318 319 // Use the original CalledValue, in case it was an alias. We want 320 // to record the call edge to the alias in that case. Eventually 321 // an alias summary will be created to associate the alias and 322 // aliasee. 323 auto &ValueInfo = CallGraphEdges[Index.getOrInsertValueInfo( 324 cast<GlobalValue>(CalledValue))]; 325 ValueInfo.updateHotness(Hotness); 326 // Add the relative block frequency to CalleeInfo if there is no profile 327 // information. 328 if (BFI != nullptr && Hotness == CalleeInfo::HotnessType::Unknown) { 329 uint64_t BBFreq = BFI->getBlockFreq(&BB).getFrequency(); 330 uint64_t EntryFreq = BFI->getEntryFreq(); 331 ValueInfo.updateRelBlockFreq(BBFreq, EntryFreq); 332 } 333 } else { 334 // Skip inline assembly calls. 335 if (CI && CI->isInlineAsm()) 336 continue; 337 // Skip direct calls. 338 if (!CalledValue || isa<Constant>(CalledValue)) 339 continue; 340 341 // Check if the instruction has a callees metadata. If so, add callees 342 // to CallGraphEdges to reflect the references from the metadata, and 343 // to enable importing for subsequent indirect call promotion and 344 // inlining. 345 if (auto *MD = I.getMetadata(LLVMContext::MD_callees)) { 346 for (auto &Op : MD->operands()) { 347 Function *Callee = mdconst::extract_or_null<Function>(Op); 348 if (Callee) 349 CallGraphEdges[Index.getOrInsertValueInfo(Callee)]; 350 } 351 } 352 353 uint32_t NumVals, NumCandidates; 354 uint64_t TotalCount; 355 auto CandidateProfileData = 356 ICallAnalysis.getPromotionCandidatesForInstruction( 357 &I, NumVals, TotalCount, NumCandidates); 358 for (auto &Candidate : CandidateProfileData) 359 CallGraphEdges[Index.getOrInsertValueInfo(Candidate.Value)] 360 .updateHotness(getHotness(Candidate.Count, PSI)); 361 } 362 } 363 364 // By now we processed all instructions in a function, except 365 // non-volatile loads. All new refs we add in a loop below 366 // are obviously constant. All constant refs are grouped in the 367 // end of RefEdges vector, so we can use a single integer value 368 // to identify them. 369 unsigned RefCnt = RefEdges.size(); 370 for (const Instruction *I : NonVolatileLoads) { 371 Visited.erase(I); 372 findRefEdges(Index, I, RefEdges, Visited); 373 } 374 std::vector<ValueInfo> Refs = RefEdges.takeVector(); 375 // Regular LTO module doesn't participate in ThinLTO import, 376 // so no reference from it can be readonly, since this would 377 // require importing variable as local copy 378 if (IsThinLTO) 379 for (; RefCnt < Refs.size(); ++RefCnt) 380 Refs[RefCnt].setReadOnly(); 381 382 // Explicit add hot edges to enforce importing for designated GUIDs for 383 // sample PGO, to enable the same inlines as the profiled optimized binary. 384 for (auto &I : F.getImportGUIDs()) 385 CallGraphEdges[Index.getOrInsertValueInfo(I)].updateHotness( 386 ForceSummaryEdgesCold == FunctionSummary::FSHT_All 387 ? CalleeInfo::HotnessType::Cold 388 : CalleeInfo::HotnessType::Critical); 389 390 bool NonRenamableLocal = isNonRenamableLocal(F); 391 bool NotEligibleForImport = 392 NonRenamableLocal || HasInlineAsmMaybeReferencingInternal; 393 GlobalValueSummary::GVFlags Flags(F.getLinkage(), NotEligibleForImport, 394 /* Live = */ false, F.isDSOLocal(), 395 F.hasLinkOnceODRLinkage() && F.hasGlobalUnnamedAddr()); 396 FunctionSummary::FFlags FunFlags{ 397 F.hasFnAttribute(Attribute::ReadNone), 398 F.hasFnAttribute(Attribute::ReadOnly), 399 F.hasFnAttribute(Attribute::NoRecurse), F.returnDoesNotAlias(), 400 // FIXME: refactor this to use the same code that inliner is using. 401 // Don't try to import functions with noinline attribute. 402 F.getAttributes().hasFnAttribute(Attribute::NoInline)}; 403 auto FuncSummary = llvm::make_unique<FunctionSummary>( 404 Flags, NumInsts, FunFlags, /*EntryCount=*/0, std::move(Refs), 405 CallGraphEdges.takeVector(), TypeTests.takeVector(), 406 TypeTestAssumeVCalls.takeVector(), TypeCheckedLoadVCalls.takeVector(), 407 TypeTestAssumeConstVCalls.takeVector(), 408 TypeCheckedLoadConstVCalls.takeVector()); 409 if (NonRenamableLocal) 410 CantBePromoted.insert(F.getGUID()); 411 Index.addGlobalValueSummary(F, std::move(FuncSummary)); 412 } 413 414 static void 415 computeVariableSummary(ModuleSummaryIndex &Index, const GlobalVariable &V, 416 DenseSet<GlobalValue::GUID> &CantBePromoted) { 417 SetVector<ValueInfo> RefEdges; 418 SmallPtrSet<const User *, 8> Visited; 419 bool HasBlockAddress = findRefEdges(Index, &V, RefEdges, Visited); 420 bool NonRenamableLocal = isNonRenamableLocal(V); 421 GlobalValueSummary::GVFlags Flags(V.getLinkage(), NonRenamableLocal, 422 /* Live = */ false, V.isDSOLocal(), 423 V.hasLinkOnceODRLinkage() && V.hasGlobalUnnamedAddr()); 424 425 // Don't mark variables we won't be able to internalize as read-only. 426 GlobalVarSummary::GVarFlags VarFlags( 427 !V.hasComdat() && !V.hasAppendingLinkage() && !V.isInterposable() && 428 !V.hasAvailableExternallyLinkage() && !V.hasDLLExportStorageClass()); 429 auto GVarSummary = llvm::make_unique<GlobalVarSummary>(Flags, VarFlags, 430 RefEdges.takeVector()); 431 if (NonRenamableLocal) 432 CantBePromoted.insert(V.getGUID()); 433 if (HasBlockAddress) 434 GVarSummary->setNotEligibleToImport(); 435 Index.addGlobalValueSummary(V, std::move(GVarSummary)); 436 } 437 438 static void 439 computeAliasSummary(ModuleSummaryIndex &Index, const GlobalAlias &A, 440 DenseSet<GlobalValue::GUID> &CantBePromoted) { 441 bool NonRenamableLocal = isNonRenamableLocal(A); 442 GlobalValueSummary::GVFlags Flags(A.getLinkage(), NonRenamableLocal, 443 /* Live = */ false, A.isDSOLocal(), 444 A.hasLinkOnceODRLinkage() && A.hasGlobalUnnamedAddr()); 445 auto AS = llvm::make_unique<AliasSummary>(Flags); 446 auto *Aliasee = A.getBaseObject(); 447 auto AliaseeVI = Index.getValueInfo(Aliasee->getGUID()); 448 assert(AliaseeVI && "Alias expects aliasee summary to be available"); 449 assert(AliaseeVI.getSummaryList().size() == 1 && 450 "Expected a single entry per aliasee in per-module index"); 451 AS->setAliasee(AliaseeVI, AliaseeVI.getSummaryList()[0].get()); 452 if (NonRenamableLocal) 453 CantBePromoted.insert(A.getGUID()); 454 Index.addGlobalValueSummary(A, std::move(AS)); 455 } 456 457 // Set LiveRoot flag on entries matching the given value name. 458 static void setLiveRoot(ModuleSummaryIndex &Index, StringRef Name) { 459 if (ValueInfo VI = Index.getValueInfo(GlobalValue::getGUID(Name))) 460 for (auto &Summary : VI.getSummaryList()) 461 Summary->setLive(true); 462 } 463 464 ModuleSummaryIndex llvm::buildModuleSummaryIndex( 465 const Module &M, 466 std::function<BlockFrequencyInfo *(const Function &F)> GetBFICallback, 467 ProfileSummaryInfo *PSI) { 468 assert(PSI); 469 bool EnableSplitLTOUnit = false; 470 if (auto *MD = mdconst::extract_or_null<ConstantInt>( 471 M.getModuleFlag("EnableSplitLTOUnit"))) 472 EnableSplitLTOUnit = MD->getZExtValue(); 473 ModuleSummaryIndex Index(/*HaveGVs=*/true, EnableSplitLTOUnit); 474 475 // Identify the local values in the llvm.used and llvm.compiler.used sets, 476 // which should not be exported as they would then require renaming and 477 // promotion, but we may have opaque uses e.g. in inline asm. We collect them 478 // here because we use this information to mark functions containing inline 479 // assembly calls as not importable. 480 SmallPtrSet<GlobalValue *, 8> LocalsUsed; 481 SmallPtrSet<GlobalValue *, 8> Used; 482 // First collect those in the llvm.used set. 483 collectUsedGlobalVariables(M, Used, /*CompilerUsed*/ false); 484 // Next collect those in the llvm.compiler.used set. 485 collectUsedGlobalVariables(M, Used, /*CompilerUsed*/ true); 486 DenseSet<GlobalValue::GUID> CantBePromoted; 487 for (auto *V : Used) { 488 if (V->hasLocalLinkage()) { 489 LocalsUsed.insert(V); 490 CantBePromoted.insert(V->getGUID()); 491 } 492 } 493 494 bool HasLocalInlineAsmSymbol = false; 495 if (!M.getModuleInlineAsm().empty()) { 496 // Collect the local values defined by module level asm, and set up 497 // summaries for these symbols so that they can be marked as NoRename, 498 // to prevent export of any use of them in regular IR that would require 499 // renaming within the module level asm. Note we don't need to create a 500 // summary for weak or global defs, as they don't need to be flagged as 501 // NoRename, and defs in module level asm can't be imported anyway. 502 // Also, any values used but not defined within module level asm should 503 // be listed on the llvm.used or llvm.compiler.used global and marked as 504 // referenced from there. 505 ModuleSymbolTable::CollectAsmSymbols( 506 M, [&](StringRef Name, object::BasicSymbolRef::Flags Flags) { 507 // Symbols not marked as Weak or Global are local definitions. 508 if (Flags & (object::BasicSymbolRef::SF_Weak | 509 object::BasicSymbolRef::SF_Global)) 510 return; 511 HasLocalInlineAsmSymbol = true; 512 GlobalValue *GV = M.getNamedValue(Name); 513 if (!GV) 514 return; 515 assert(GV->isDeclaration() && "Def in module asm already has definition"); 516 GlobalValueSummary::GVFlags GVFlags(GlobalValue::InternalLinkage, 517 /* NotEligibleToImport = */ true, 518 /* Live = */ true, 519 /* Local */ GV->isDSOLocal(), 520 GV->hasLinkOnceODRLinkage() && GV->hasGlobalUnnamedAddr()); 521 CantBePromoted.insert(GV->getGUID()); 522 // Create the appropriate summary type. 523 if (Function *F = dyn_cast<Function>(GV)) { 524 std::unique_ptr<FunctionSummary> Summary = 525 llvm::make_unique<FunctionSummary>( 526 GVFlags, /*InstCount=*/0, 527 FunctionSummary::FFlags{ 528 F->hasFnAttribute(Attribute::ReadNone), 529 F->hasFnAttribute(Attribute::ReadOnly), 530 F->hasFnAttribute(Attribute::NoRecurse), 531 F->returnDoesNotAlias(), 532 /* NoInline = */ false}, 533 /*EntryCount=*/0, ArrayRef<ValueInfo>{}, 534 ArrayRef<FunctionSummary::EdgeTy>{}, 535 ArrayRef<GlobalValue::GUID>{}, 536 ArrayRef<FunctionSummary::VFuncId>{}, 537 ArrayRef<FunctionSummary::VFuncId>{}, 538 ArrayRef<FunctionSummary::ConstVCall>{}, 539 ArrayRef<FunctionSummary::ConstVCall>{}); 540 Index.addGlobalValueSummary(*GV, std::move(Summary)); 541 } else { 542 std::unique_ptr<GlobalVarSummary> Summary = 543 llvm::make_unique<GlobalVarSummary>( 544 GVFlags, GlobalVarSummary::GVarFlags(), 545 ArrayRef<ValueInfo>{}); 546 Index.addGlobalValueSummary(*GV, std::move(Summary)); 547 } 548 }); 549 } 550 551 bool IsThinLTO = true; 552 if (auto *MD = 553 mdconst::extract_or_null<ConstantInt>(M.getModuleFlag("ThinLTO"))) 554 IsThinLTO = MD->getZExtValue(); 555 556 // Compute summaries for all functions defined in module, and save in the 557 // index. 558 for (auto &F : M) { 559 if (F.isDeclaration()) 560 continue; 561 562 DominatorTree DT(const_cast<Function &>(F)); 563 BlockFrequencyInfo *BFI = nullptr; 564 std::unique_ptr<BlockFrequencyInfo> BFIPtr; 565 if (GetBFICallback) 566 BFI = GetBFICallback(F); 567 else if (F.hasProfileData()) { 568 LoopInfo LI{DT}; 569 BranchProbabilityInfo BPI{F, LI}; 570 BFIPtr = llvm::make_unique<BlockFrequencyInfo>(F, BPI, LI); 571 BFI = BFIPtr.get(); 572 } 573 574 computeFunctionSummary(Index, M, F, BFI, PSI, DT, 575 !LocalsUsed.empty() || HasLocalInlineAsmSymbol, 576 CantBePromoted, IsThinLTO); 577 } 578 579 // Compute summaries for all variables defined in module, and save in the 580 // index. 581 for (const GlobalVariable &G : M.globals()) { 582 if (G.isDeclaration()) 583 continue; 584 computeVariableSummary(Index, G, CantBePromoted); 585 } 586 587 // Compute summaries for all aliases defined in module, and save in the 588 // index. 589 for (const GlobalAlias &A : M.aliases()) 590 computeAliasSummary(Index, A, CantBePromoted); 591 592 for (auto *V : LocalsUsed) { 593 auto *Summary = Index.getGlobalValueSummary(*V); 594 assert(Summary && "Missing summary for global value"); 595 Summary->setNotEligibleToImport(); 596 } 597 598 // The linker doesn't know about these LLVM produced values, so we need 599 // to flag them as live in the index to ensure index-based dead value 600 // analysis treats them as live roots of the analysis. 601 setLiveRoot(Index, "llvm.used"); 602 setLiveRoot(Index, "llvm.compiler.used"); 603 setLiveRoot(Index, "llvm.global_ctors"); 604 setLiveRoot(Index, "llvm.global_dtors"); 605 setLiveRoot(Index, "llvm.global.annotations"); 606 607 for (auto &GlobalList : Index) { 608 // Ignore entries for references that are undefined in the current module. 609 if (GlobalList.second.SummaryList.empty()) 610 continue; 611 612 assert(GlobalList.second.SummaryList.size() == 1 && 613 "Expected module's index to have one summary per GUID"); 614 auto &Summary = GlobalList.second.SummaryList[0]; 615 if (!IsThinLTO) { 616 Summary->setNotEligibleToImport(); 617 continue; 618 } 619 620 bool AllRefsCanBeExternallyReferenced = 621 llvm::all_of(Summary->refs(), [&](const ValueInfo &VI) { 622 return !CantBePromoted.count(VI.getGUID()); 623 }); 624 if (!AllRefsCanBeExternallyReferenced) { 625 Summary->setNotEligibleToImport(); 626 continue; 627 } 628 629 if (auto *FuncSummary = dyn_cast<FunctionSummary>(Summary.get())) { 630 bool AllCallsCanBeExternallyReferenced = llvm::all_of( 631 FuncSummary->calls(), [&](const FunctionSummary::EdgeTy &Edge) { 632 return !CantBePromoted.count(Edge.first.getGUID()); 633 }); 634 if (!AllCallsCanBeExternallyReferenced) 635 Summary->setNotEligibleToImport(); 636 } 637 } 638 639 if (!ModuleSummaryDotFile.empty()) { 640 std::error_code EC; 641 raw_fd_ostream OSDot(ModuleSummaryDotFile, EC, sys::fs::OpenFlags::F_None); 642 if (EC) 643 report_fatal_error(Twine("Failed to open dot file ") + 644 ModuleSummaryDotFile + ": " + EC.message() + "\n"); 645 Index.exportToDot(OSDot); 646 } 647 648 return Index; 649 } 650 651 AnalysisKey ModuleSummaryIndexAnalysis::Key; 652 653 ModuleSummaryIndex 654 ModuleSummaryIndexAnalysis::run(Module &M, ModuleAnalysisManager &AM) { 655 ProfileSummaryInfo &PSI = AM.getResult<ProfileSummaryAnalysis>(M); 656 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 657 return buildModuleSummaryIndex( 658 M, 659 [&FAM](const Function &F) { 660 return &FAM.getResult<BlockFrequencyAnalysis>( 661 *const_cast<Function *>(&F)); 662 }, 663 &PSI); 664 } 665 666 char ModuleSummaryIndexWrapperPass::ID = 0; 667 668 INITIALIZE_PASS_BEGIN(ModuleSummaryIndexWrapperPass, "module-summary-analysis", 669 "Module Summary Analysis", false, true) 670 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) 671 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass) 672 INITIALIZE_PASS_END(ModuleSummaryIndexWrapperPass, "module-summary-analysis", 673 "Module Summary Analysis", false, true) 674 675 ModulePass *llvm::createModuleSummaryIndexWrapperPass() { 676 return new ModuleSummaryIndexWrapperPass(); 677 } 678 679 ModuleSummaryIndexWrapperPass::ModuleSummaryIndexWrapperPass() 680 : ModulePass(ID) { 681 initializeModuleSummaryIndexWrapperPassPass(*PassRegistry::getPassRegistry()); 682 } 683 684 bool ModuleSummaryIndexWrapperPass::runOnModule(Module &M) { 685 auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 686 Index.emplace(buildModuleSummaryIndex( 687 M, 688 [this](const Function &F) { 689 return &(this->getAnalysis<BlockFrequencyInfoWrapperPass>( 690 *const_cast<Function *>(&F)) 691 .getBFI()); 692 }, 693 PSI)); 694 return false; 695 } 696 697 bool ModuleSummaryIndexWrapperPass::doFinalization(Module &M) { 698 Index.reset(); 699 return false; 700 } 701 702 void ModuleSummaryIndexWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 703 AU.setPreservesAll(); 704 AU.addRequired<BlockFrequencyInfoWrapperPass>(); 705 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 706 } 707