1 //===- WholeProgramDevirt.cpp - Whole program virtual call optimization ---===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This pass implements whole program optimization of virtual calls in cases 11 // where we know (via !type metadata) that the list of callees is fixed. This 12 // includes the following: 13 // - Single implementation devirtualization: if a virtual call has a single 14 // possible callee, replace all calls with a direct call to that callee. 15 // - Virtual constant propagation: if the virtual function's return type is an 16 // integer <=64 bits and all possible callees are readnone, for each class and 17 // each list of constant arguments: evaluate the function, store the return 18 // value alongside the virtual table, and rewrite each virtual call as a load 19 // from the virtual table. 20 // - Uniform return value optimization: if the conditions for virtual constant 21 // propagation hold and each function returns the same constant value, replace 22 // each virtual call with that constant. 23 // - Unique return value optimization for i1 return values: if the conditions 24 // for virtual constant propagation hold and a single vtable's function 25 // returns 0, or a single vtable's function returns 1, replace each virtual 26 // call with a comparison of the vptr against that vtable's address. 27 // 28 // This pass is intended to be used during the regular and thin LTO pipelines. 29 // During regular LTO, the pass determines the best optimization for each 30 // virtual call and applies the resolutions directly to virtual calls that are 31 // eligible for virtual call optimization (i.e. calls that use either of the 32 // llvm.assume(llvm.type.test) or llvm.type.checked.load intrinsics). During 33 // ThinLTO, the pass operates in two phases: 34 // - Export phase: this is run during the thin link over a single merged module 35 // that contains all vtables with !type metadata that participate in the link. 36 // The pass computes a resolution for each virtual call and stores it in the 37 // type identifier summary. 38 // - Import phase: this is run during the thin backends over the individual 39 // modules. The pass applies the resolutions previously computed during the 40 // import phase to each eligible virtual call. 41 // 42 //===----------------------------------------------------------------------===// 43 44 #include "llvm/Transforms/IPO/WholeProgramDevirt.h" 45 #include "llvm/ADT/ArrayRef.h" 46 #include "llvm/ADT/DenseMap.h" 47 #include "llvm/ADT/DenseMapInfo.h" 48 #include "llvm/ADT/DenseSet.h" 49 #include "llvm/ADT/MapVector.h" 50 #include "llvm/ADT/SmallVector.h" 51 #include "llvm/ADT/iterator_range.h" 52 #include "llvm/Analysis/AliasAnalysis.h" 53 #include "llvm/Analysis/BasicAliasAnalysis.h" 54 #include "llvm/Analysis/OptimizationDiagnosticInfo.h" 55 #include "llvm/Analysis/TypeMetadataUtils.h" 56 #include "llvm/IR/CallSite.h" 57 #include "llvm/IR/Constants.h" 58 #include "llvm/IR/DataLayout.h" 59 #include "llvm/IR/DebugInfoMetadata.h" 60 #include "llvm/IR/DebugLoc.h" 61 #include "llvm/IR/DerivedTypes.h" 62 #include "llvm/IR/Function.h" 63 #include "llvm/IR/GlobalAlias.h" 64 #include "llvm/IR/GlobalVariable.h" 65 #include "llvm/IR/IRBuilder.h" 66 #include "llvm/IR/InstrTypes.h" 67 #include "llvm/IR/Instruction.h" 68 #include "llvm/IR/Instructions.h" 69 #include "llvm/IR/Intrinsics.h" 70 #include "llvm/IR/LLVMContext.h" 71 #include "llvm/IR/Metadata.h" 72 #include "llvm/IR/Module.h" 73 #include "llvm/IR/ModuleSummaryIndexYAML.h" 74 #include "llvm/Pass.h" 75 #include "llvm/PassRegistry.h" 76 #include "llvm/PassSupport.h" 77 #include "llvm/Support/Casting.h" 78 #include "llvm/Support/Error.h" 79 #include "llvm/Support/FileSystem.h" 80 #include "llvm/Support/MathExtras.h" 81 #include "llvm/Transforms/IPO.h" 82 #include "llvm/Transforms/IPO/FunctionAttrs.h" 83 #include "llvm/Transforms/Utils/Evaluator.h" 84 #include <algorithm> 85 #include <cstddef> 86 #include <map> 87 #include <set> 88 #include <string> 89 90 using namespace llvm; 91 using namespace wholeprogramdevirt; 92 93 #define DEBUG_TYPE "wholeprogramdevirt" 94 95 static cl::opt<PassSummaryAction> ClSummaryAction( 96 "wholeprogramdevirt-summary-action", 97 cl::desc("What to do with the summary when running this pass"), 98 cl::values(clEnumValN(PassSummaryAction::None, "none", "Do nothing"), 99 clEnumValN(PassSummaryAction::Import, "import", 100 "Import typeid resolutions from summary and globals"), 101 clEnumValN(PassSummaryAction::Export, "export", 102 "Export typeid resolutions to summary and globals")), 103 cl::Hidden); 104 105 static cl::opt<std::string> ClReadSummary( 106 "wholeprogramdevirt-read-summary", 107 cl::desc("Read summary from given YAML file before running pass"), 108 cl::Hidden); 109 110 static cl::opt<std::string> ClWriteSummary( 111 "wholeprogramdevirt-write-summary", 112 cl::desc("Write summary to given YAML file after running pass"), 113 cl::Hidden); 114 115 // Find the minimum offset that we may store a value of size Size bits at. If 116 // IsAfter is set, look for an offset before the object, otherwise look for an 117 // offset after the object. 118 uint64_t 119 wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets, 120 bool IsAfter, uint64_t Size) { 121 // Find a minimum offset taking into account only vtable sizes. 122 uint64_t MinByte = 0; 123 for (const VirtualCallTarget &Target : Targets) { 124 if (IsAfter) 125 MinByte = std::max(MinByte, Target.minAfterBytes()); 126 else 127 MinByte = std::max(MinByte, Target.minBeforeBytes()); 128 } 129 130 // Build a vector of arrays of bytes covering, for each target, a slice of the 131 // used region (see AccumBitVector::BytesUsed in 132 // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively, 133 // this aligns the used regions to start at MinByte. 134 // 135 // In this example, A, B and C are vtables, # is a byte already allocated for 136 // a virtual function pointer, AAAA... (etc.) are the used regions for the 137 // vtables and Offset(X) is the value computed for the Offset variable below 138 // for X. 139 // 140 // Offset(A) 141 // | | 142 // |MinByte 143 // A: ################AAAAAAAA|AAAAAAAA 144 // B: ########BBBBBBBBBBBBBBBB|BBBB 145 // C: ########################|CCCCCCCCCCCCCCCC 146 // | Offset(B) | 147 // 148 // This code produces the slices of A, B and C that appear after the divider 149 // at MinByte. 150 std::vector<ArrayRef<uint8_t>> Used; 151 for (const VirtualCallTarget &Target : Targets) { 152 ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed 153 : Target.TM->Bits->Before.BytesUsed; 154 uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes() 155 : MinByte - Target.minBeforeBytes(); 156 157 // Disregard used regions that are smaller than Offset. These are 158 // effectively all-free regions that do not need to be checked. 159 if (VTUsed.size() > Offset) 160 Used.push_back(VTUsed.slice(Offset)); 161 } 162 163 if (Size == 1) { 164 // Find a free bit in each member of Used. 165 for (unsigned I = 0;; ++I) { 166 uint8_t BitsUsed = 0; 167 for (auto &&B : Used) 168 if (I < B.size()) 169 BitsUsed |= B[I]; 170 if (BitsUsed != 0xff) 171 return (MinByte + I) * 8 + 172 countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined); 173 } 174 } else { 175 // Find a free (Size/8) byte region in each member of Used. 176 // FIXME: see if alignment helps. 177 for (unsigned I = 0;; ++I) { 178 for (auto &&B : Used) { 179 unsigned Byte = 0; 180 while ((I + Byte) < B.size() && Byte < (Size / 8)) { 181 if (B[I + Byte]) 182 goto NextI; 183 ++Byte; 184 } 185 } 186 return (MinByte + I) * 8; 187 NextI:; 188 } 189 } 190 } 191 192 void wholeprogramdevirt::setBeforeReturnValues( 193 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore, 194 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 195 if (BitWidth == 1) 196 OffsetByte = -(AllocBefore / 8 + 1); 197 else 198 OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8); 199 OffsetBit = AllocBefore % 8; 200 201 for (VirtualCallTarget &Target : Targets) { 202 if (BitWidth == 1) 203 Target.setBeforeBit(AllocBefore); 204 else 205 Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8); 206 } 207 } 208 209 void wholeprogramdevirt::setAfterReturnValues( 210 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter, 211 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 212 if (BitWidth == 1) 213 OffsetByte = AllocAfter / 8; 214 else 215 OffsetByte = (AllocAfter + 7) / 8; 216 OffsetBit = AllocAfter % 8; 217 218 for (VirtualCallTarget &Target : Targets) { 219 if (BitWidth == 1) 220 Target.setAfterBit(AllocAfter); 221 else 222 Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8); 223 } 224 } 225 226 VirtualCallTarget::VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM) 227 : Fn(Fn), TM(TM), 228 IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {} 229 230 namespace { 231 232 // A slot in a set of virtual tables. The TypeID identifies the set of virtual 233 // tables, and the ByteOffset is the offset in bytes from the address point to 234 // the virtual function pointer. 235 struct VTableSlot { 236 Metadata *TypeID; 237 uint64_t ByteOffset; 238 }; 239 240 } // end anonymous namespace 241 242 namespace llvm { 243 244 template <> struct DenseMapInfo<VTableSlot> { 245 static VTableSlot getEmptyKey() { 246 return {DenseMapInfo<Metadata *>::getEmptyKey(), 247 DenseMapInfo<uint64_t>::getEmptyKey()}; 248 } 249 static VTableSlot getTombstoneKey() { 250 return {DenseMapInfo<Metadata *>::getTombstoneKey(), 251 DenseMapInfo<uint64_t>::getTombstoneKey()}; 252 } 253 static unsigned getHashValue(const VTableSlot &I) { 254 return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^ 255 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 256 } 257 static bool isEqual(const VTableSlot &LHS, 258 const VTableSlot &RHS) { 259 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 260 } 261 }; 262 263 } // end namespace llvm 264 265 namespace { 266 267 // A virtual call site. VTable is the loaded virtual table pointer, and CS is 268 // the indirect virtual call. 269 struct VirtualCallSite { 270 Value *VTable; 271 CallSite CS; 272 273 // If non-null, this field points to the associated unsafe use count stored in 274 // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description 275 // of that field for details. 276 unsigned *NumUnsafeUses; 277 278 void 279 emitRemark(const StringRef OptName, const StringRef TargetName, 280 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) { 281 Function *F = CS.getCaller(); 282 DebugLoc DLoc = CS->getDebugLoc(); 283 BasicBlock *Block = CS.getParent(); 284 285 // In the new pass manager, we can request the optimization 286 // remark emitter pass on a per-function-basis, which the 287 // OREGetter will do for us. 288 // In the old pass manager, this is harder, so we just build 289 // a optimization remark emitter on the fly, when we need it. 290 std::unique_ptr<OptimizationRemarkEmitter> OwnedORE; 291 OptimizationRemarkEmitter *ORE; 292 if (OREGetter) 293 ORE = &OREGetter(F); 294 else { 295 OwnedORE = make_unique<OptimizationRemarkEmitter>(F); 296 ORE = OwnedORE.get(); 297 } 298 299 using namespace ore; 300 ORE->emit(OptimizationRemark(DEBUG_TYPE, OptName, DLoc, Block) 301 << NV("Optimization", OptName) << ": devirtualized a call to " 302 << NV("FunctionName", TargetName)); 303 } 304 305 void replaceAndErase( 306 const StringRef OptName, const StringRef TargetName, bool RemarksEnabled, 307 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 308 Value *New) { 309 if (RemarksEnabled) 310 emitRemark(OptName, TargetName, OREGetter); 311 CS->replaceAllUsesWith(New); 312 if (auto II = dyn_cast<InvokeInst>(CS.getInstruction())) { 313 BranchInst::Create(II->getNormalDest(), CS.getInstruction()); 314 II->getUnwindDest()->removePredecessor(II->getParent()); 315 } 316 CS->eraseFromParent(); 317 // This use is no longer unsafe. 318 if (NumUnsafeUses) 319 --*NumUnsafeUses; 320 } 321 }; 322 323 // Call site information collected for a specific VTableSlot and possibly a list 324 // of constant integer arguments. The grouping by arguments is handled by the 325 // VTableSlotInfo class. 326 struct CallSiteInfo { 327 /// The set of call sites for this slot. Used during regular LTO and the 328 /// import phase of ThinLTO (as well as the export phase of ThinLTO for any 329 /// call sites that appear in the merged module itself); in each of these 330 /// cases we are directly operating on the call sites at the IR level. 331 std::vector<VirtualCallSite> CallSites; 332 333 // These fields are used during the export phase of ThinLTO and reflect 334 // information collected from function summaries. 335 336 /// Whether any function summary contains an llvm.assume(llvm.type.test) for 337 /// this slot. 338 bool SummaryHasTypeTestAssumeUsers; 339 340 /// CFI-specific: a vector containing the list of function summaries that use 341 /// the llvm.type.checked.load intrinsic and therefore will require 342 /// resolutions for llvm.type.test in order to implement CFI checks if 343 /// devirtualization was unsuccessful. If devirtualization was successful, the 344 /// pass will clear this vector by calling markDevirt(). If at the end of the 345 /// pass the vector is non-empty, we will need to add a use of llvm.type.test 346 /// to each of the function summaries in the vector. 347 std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers; 348 349 bool isExported() const { 350 return SummaryHasTypeTestAssumeUsers || 351 !SummaryTypeCheckedLoadUsers.empty(); 352 } 353 354 /// As explained in the comment for SummaryTypeCheckedLoadUsers. 355 void markDevirt() { SummaryTypeCheckedLoadUsers.clear(); } 356 }; 357 358 // Call site information collected for a specific VTableSlot. 359 struct VTableSlotInfo { 360 // The set of call sites which do not have all constant integer arguments 361 // (excluding "this"). 362 CallSiteInfo CSInfo; 363 364 // The set of call sites with all constant integer arguments (excluding 365 // "this"), grouped by argument list. 366 std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo; 367 368 void addCallSite(Value *VTable, CallSite CS, unsigned *NumUnsafeUses); 369 370 private: 371 CallSiteInfo &findCallSiteInfo(CallSite CS); 372 }; 373 374 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallSite CS) { 375 std::vector<uint64_t> Args; 376 auto *CI = dyn_cast<IntegerType>(CS.getType()); 377 if (!CI || CI->getBitWidth() > 64 || CS.arg_empty()) 378 return CSInfo; 379 for (auto &&Arg : make_range(CS.arg_begin() + 1, CS.arg_end())) { 380 auto *CI = dyn_cast<ConstantInt>(Arg); 381 if (!CI || CI->getBitWidth() > 64) 382 return CSInfo; 383 Args.push_back(CI->getZExtValue()); 384 } 385 return ConstCSInfo[Args]; 386 } 387 388 void VTableSlotInfo::addCallSite(Value *VTable, CallSite CS, 389 unsigned *NumUnsafeUses) { 390 findCallSiteInfo(CS).CallSites.push_back({VTable, CS, NumUnsafeUses}); 391 } 392 393 struct DevirtModule { 394 Module &M; 395 function_ref<AAResults &(Function &)> AARGetter; 396 397 ModuleSummaryIndex *ExportSummary; 398 const ModuleSummaryIndex *ImportSummary; 399 400 IntegerType *Int8Ty; 401 PointerType *Int8PtrTy; 402 IntegerType *Int32Ty; 403 IntegerType *Int64Ty; 404 IntegerType *IntPtrTy; 405 406 bool RemarksEnabled; 407 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter; 408 409 MapVector<VTableSlot, VTableSlotInfo> CallSlots; 410 411 // This map keeps track of the number of "unsafe" uses of a loaded function 412 // pointer. The key is the associated llvm.type.test intrinsic call generated 413 // by this pass. An unsafe use is one that calls the loaded function pointer 414 // directly. Every time we eliminate an unsafe use (for example, by 415 // devirtualizing it or by applying virtual constant propagation), we 416 // decrement the value stored in this map. If a value reaches zero, we can 417 // eliminate the type check by RAUWing the associated llvm.type.test call with 418 // true. 419 std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest; 420 421 DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter, 422 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 423 ModuleSummaryIndex *ExportSummary, 424 const ModuleSummaryIndex *ImportSummary) 425 : M(M), AARGetter(AARGetter), ExportSummary(ExportSummary), 426 ImportSummary(ImportSummary), Int8Ty(Type::getInt8Ty(M.getContext())), 427 Int8PtrTy(Type::getInt8PtrTy(M.getContext())), 428 Int32Ty(Type::getInt32Ty(M.getContext())), 429 Int64Ty(Type::getInt64Ty(M.getContext())), 430 IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)), 431 RemarksEnabled(areRemarksEnabled()), OREGetter(OREGetter) { 432 assert(!(ExportSummary && ImportSummary)); 433 } 434 435 bool areRemarksEnabled(); 436 437 void scanTypeTestUsers(Function *TypeTestFunc, Function *AssumeFunc); 438 void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc); 439 440 void buildTypeIdentifierMap( 441 std::vector<VTableBits> &Bits, 442 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap); 443 Constant *getPointerAtOffset(Constant *I, uint64_t Offset); 444 bool 445 tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot, 446 const std::set<TypeMemberInfo> &TypeMemberInfos, 447 uint64_t ByteOffset); 448 449 void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn, 450 bool &IsExported); 451 bool trySingleImplDevirt(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 452 VTableSlotInfo &SlotInfo, 453 WholeProgramDevirtResolution *Res); 454 455 bool tryEvaluateFunctionsWithArgs( 456 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 457 ArrayRef<uint64_t> Args); 458 459 void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 460 uint64_t TheRetVal); 461 bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 462 CallSiteInfo &CSInfo, 463 WholeProgramDevirtResolution::ByArg *Res); 464 465 // Returns the global symbol name that is used to export information about the 466 // given vtable slot and list of arguments. 467 std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args, 468 StringRef Name); 469 470 // This function is called during the export phase to create a symbol 471 // definition containing information about the given vtable slot and list of 472 // arguments. 473 void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 474 Constant *C); 475 476 // This function is called during the import phase to create a reference to 477 // the symbol definition created during the export phase. 478 Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 479 StringRef Name, unsigned AbsWidth = 0); 480 481 void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne, 482 Constant *UniqueMemberAddr); 483 bool tryUniqueRetValOpt(unsigned BitWidth, 484 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 485 CallSiteInfo &CSInfo, 486 WholeProgramDevirtResolution::ByArg *Res, 487 VTableSlot Slot, ArrayRef<uint64_t> Args); 488 489 void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 490 Constant *Byte, Constant *Bit); 491 bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 492 VTableSlotInfo &SlotInfo, 493 WholeProgramDevirtResolution *Res, VTableSlot Slot); 494 495 void rebuildGlobal(VTableBits &B); 496 497 // Apply the summary resolution for Slot to all virtual calls in SlotInfo. 498 void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo); 499 500 // If we were able to eliminate all unsafe uses for a type checked load, 501 // eliminate the associated type tests by replacing them with true. 502 void removeRedundantTypeTests(); 503 504 bool run(); 505 506 // Lower the module using the action and summary passed as command line 507 // arguments. For testing purposes only. 508 static bool runForTesting( 509 Module &M, function_ref<AAResults &(Function &)> AARGetter, 510 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter); 511 }; 512 513 struct WholeProgramDevirt : public ModulePass { 514 static char ID; 515 516 bool UseCommandLine = false; 517 518 ModuleSummaryIndex *ExportSummary; 519 const ModuleSummaryIndex *ImportSummary; 520 521 WholeProgramDevirt() : ModulePass(ID), UseCommandLine(true) { 522 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry()); 523 } 524 525 WholeProgramDevirt(ModuleSummaryIndex *ExportSummary, 526 const ModuleSummaryIndex *ImportSummary) 527 : ModulePass(ID), ExportSummary(ExportSummary), 528 ImportSummary(ImportSummary) { 529 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry()); 530 } 531 532 bool runOnModule(Module &M) override { 533 if (skipModule(M)) 534 return false; 535 536 auto OREGetter = function_ref<OptimizationRemarkEmitter &(Function *)>(); 537 538 if (UseCommandLine) 539 return DevirtModule::runForTesting(M, LegacyAARGetter(*this), OREGetter); 540 541 return DevirtModule(M, LegacyAARGetter(*this), OREGetter, ExportSummary, 542 ImportSummary) 543 .run(); 544 } 545 546 void getAnalysisUsage(AnalysisUsage &AU) const override { 547 AU.addRequired<AssumptionCacheTracker>(); 548 AU.addRequired<TargetLibraryInfoWrapperPass>(); 549 } 550 }; 551 552 } // end anonymous namespace 553 554 INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt", 555 "Whole program devirtualization", false, false) 556 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 557 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 558 INITIALIZE_PASS_END(WholeProgramDevirt, "wholeprogramdevirt", 559 "Whole program devirtualization", false, false) 560 char WholeProgramDevirt::ID = 0; 561 562 ModulePass * 563 llvm::createWholeProgramDevirtPass(ModuleSummaryIndex *ExportSummary, 564 const ModuleSummaryIndex *ImportSummary) { 565 return new WholeProgramDevirt(ExportSummary, ImportSummary); 566 } 567 568 PreservedAnalyses WholeProgramDevirtPass::run(Module &M, 569 ModuleAnalysisManager &AM) { 570 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 571 auto AARGetter = [&](Function &F) -> AAResults & { 572 return FAM.getResult<AAManager>(F); 573 }; 574 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { 575 return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F); 576 }; 577 if (!DevirtModule(M, AARGetter, OREGetter, nullptr, nullptr).run()) 578 return PreservedAnalyses::all(); 579 return PreservedAnalyses::none(); 580 } 581 582 bool DevirtModule::runForTesting( 583 Module &M, function_ref<AAResults &(Function &)> AARGetter, 584 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) { 585 ModuleSummaryIndex Summary; 586 587 // Handle the command-line summary arguments. This code is for testing 588 // purposes only, so we handle errors directly. 589 if (!ClReadSummary.empty()) { 590 ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary + 591 ": "); 592 auto ReadSummaryFile = 593 ExitOnErr(errorOrToExpected(MemoryBuffer::getFile(ClReadSummary))); 594 595 yaml::Input In(ReadSummaryFile->getBuffer()); 596 In >> Summary; 597 ExitOnErr(errorCodeToError(In.error())); 598 } 599 600 bool Changed = 601 DevirtModule( 602 M, AARGetter, OREGetter, 603 ClSummaryAction == PassSummaryAction::Export ? &Summary : nullptr, 604 ClSummaryAction == PassSummaryAction::Import ? &Summary : nullptr) 605 .run(); 606 607 if (!ClWriteSummary.empty()) { 608 ExitOnError ExitOnErr( 609 "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": "); 610 std::error_code EC; 611 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::F_Text); 612 ExitOnErr(errorCodeToError(EC)); 613 614 yaml::Output Out(OS); 615 Out << Summary; 616 } 617 618 return Changed; 619 } 620 621 void DevirtModule::buildTypeIdentifierMap( 622 std::vector<VTableBits> &Bits, 623 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 624 DenseMap<GlobalVariable *, VTableBits *> GVToBits; 625 Bits.reserve(M.getGlobalList().size()); 626 SmallVector<MDNode *, 2> Types; 627 for (GlobalVariable &GV : M.globals()) { 628 Types.clear(); 629 GV.getMetadata(LLVMContext::MD_type, Types); 630 if (Types.empty()) 631 continue; 632 633 VTableBits *&BitsPtr = GVToBits[&GV]; 634 if (!BitsPtr) { 635 Bits.emplace_back(); 636 Bits.back().GV = &GV; 637 Bits.back().ObjectSize = 638 M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType()); 639 BitsPtr = &Bits.back(); 640 } 641 642 for (MDNode *Type : Types) { 643 auto TypeID = Type->getOperand(1).get(); 644 645 uint64_t Offset = 646 cast<ConstantInt>( 647 cast<ConstantAsMetadata>(Type->getOperand(0))->getValue()) 648 ->getZExtValue(); 649 650 TypeIdMap[TypeID].insert({BitsPtr, Offset}); 651 } 652 } 653 } 654 655 Constant *DevirtModule::getPointerAtOffset(Constant *I, uint64_t Offset) { 656 if (I->getType()->isPointerTy()) { 657 if (Offset == 0) 658 return I; 659 return nullptr; 660 } 661 662 const DataLayout &DL = M.getDataLayout(); 663 664 if (auto *C = dyn_cast<ConstantStruct>(I)) { 665 const StructLayout *SL = DL.getStructLayout(C->getType()); 666 if (Offset >= SL->getSizeInBytes()) 667 return nullptr; 668 669 unsigned Op = SL->getElementContainingOffset(Offset); 670 return getPointerAtOffset(cast<Constant>(I->getOperand(Op)), 671 Offset - SL->getElementOffset(Op)); 672 } 673 if (auto *C = dyn_cast<ConstantArray>(I)) { 674 ArrayType *VTableTy = C->getType(); 675 uint64_t ElemSize = DL.getTypeAllocSize(VTableTy->getElementType()); 676 677 unsigned Op = Offset / ElemSize; 678 if (Op >= C->getNumOperands()) 679 return nullptr; 680 681 return getPointerAtOffset(cast<Constant>(I->getOperand(Op)), 682 Offset % ElemSize); 683 } 684 return nullptr; 685 } 686 687 bool DevirtModule::tryFindVirtualCallTargets( 688 std::vector<VirtualCallTarget> &TargetsForSlot, 689 const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset) { 690 for (const TypeMemberInfo &TM : TypeMemberInfos) { 691 if (!TM.Bits->GV->isConstant()) 692 return false; 693 694 Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(), 695 TM.Offset + ByteOffset); 696 if (!Ptr) 697 return false; 698 699 auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts()); 700 if (!Fn) 701 return false; 702 703 // We can disregard __cxa_pure_virtual as a possible call target, as 704 // calls to pure virtuals are UB. 705 if (Fn->getName() == "__cxa_pure_virtual") 706 continue; 707 708 TargetsForSlot.push_back({Fn, &TM}); 709 } 710 711 // Give up if we couldn't find any targets. 712 return !TargetsForSlot.empty(); 713 } 714 715 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo, 716 Constant *TheFn, bool &IsExported) { 717 auto Apply = [&](CallSiteInfo &CSInfo) { 718 for (auto &&VCallSite : CSInfo.CallSites) { 719 if (RemarksEnabled) 720 VCallSite.emitRemark("single-impl", TheFn->getName(), OREGetter); 721 VCallSite.CS.setCalledFunction(ConstantExpr::getBitCast( 722 TheFn, VCallSite.CS.getCalledValue()->getType())); 723 // This use is no longer unsafe. 724 if (VCallSite.NumUnsafeUses) 725 --*VCallSite.NumUnsafeUses; 726 } 727 if (CSInfo.isExported()) { 728 IsExported = true; 729 CSInfo.markDevirt(); 730 } 731 }; 732 Apply(SlotInfo.CSInfo); 733 for (auto &P : SlotInfo.ConstCSInfo) 734 Apply(P.second); 735 } 736 737 bool DevirtModule::trySingleImplDevirt( 738 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 739 VTableSlotInfo &SlotInfo, WholeProgramDevirtResolution *Res) { 740 // See if the program contains a single implementation of this virtual 741 // function. 742 Function *TheFn = TargetsForSlot[0].Fn; 743 for (auto &&Target : TargetsForSlot) 744 if (TheFn != Target.Fn) 745 return false; 746 747 // If so, update each call site to call that implementation directly. 748 if (RemarksEnabled) 749 TargetsForSlot[0].WasDevirt = true; 750 751 bool IsExported = false; 752 applySingleImplDevirt(SlotInfo, TheFn, IsExported); 753 if (!IsExported) 754 return false; 755 756 // If the only implementation has local linkage, we must promote to external 757 // to make it visible to thin LTO objects. We can only get here during the 758 // ThinLTO export phase. 759 if (TheFn->hasLocalLinkage()) { 760 std::string NewName = (TheFn->getName() + "$merged").str(); 761 762 // Since we are renaming the function, any comdats with the same name must 763 // also be renamed. This is required when targeting COFF, as the comdat name 764 // must match one of the names of the symbols in the comdat. 765 if (Comdat *C = TheFn->getComdat()) { 766 if (C->getName() == TheFn->getName()) { 767 Comdat *NewC = M.getOrInsertComdat(NewName); 768 NewC->setSelectionKind(C->getSelectionKind()); 769 for (GlobalObject &GO : M.global_objects()) 770 if (GO.getComdat() == C) 771 GO.setComdat(NewC); 772 } 773 } 774 775 TheFn->setLinkage(GlobalValue::ExternalLinkage); 776 TheFn->setVisibility(GlobalValue::HiddenVisibility); 777 TheFn->setName(NewName); 778 } 779 780 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 781 Res->SingleImplName = TheFn->getName(); 782 783 return true; 784 } 785 786 bool DevirtModule::tryEvaluateFunctionsWithArgs( 787 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 788 ArrayRef<uint64_t> Args) { 789 // Evaluate each function and store the result in each target's RetVal 790 // field. 791 for (VirtualCallTarget &Target : TargetsForSlot) { 792 if (Target.Fn->arg_size() != Args.size() + 1) 793 return false; 794 795 Evaluator Eval(M.getDataLayout(), nullptr); 796 SmallVector<Constant *, 2> EvalArgs; 797 EvalArgs.push_back( 798 Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0))); 799 for (unsigned I = 0; I != Args.size(); ++I) { 800 auto *ArgTy = dyn_cast<IntegerType>( 801 Target.Fn->getFunctionType()->getParamType(I + 1)); 802 if (!ArgTy) 803 return false; 804 EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I])); 805 } 806 807 Constant *RetVal; 808 if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) || 809 !isa<ConstantInt>(RetVal)) 810 return false; 811 Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue(); 812 } 813 return true; 814 } 815 816 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 817 uint64_t TheRetVal) { 818 for (auto Call : CSInfo.CallSites) 819 Call.replaceAndErase( 820 "uniform-ret-val", FnName, RemarksEnabled, OREGetter, 821 ConstantInt::get(cast<IntegerType>(Call.CS.getType()), TheRetVal)); 822 CSInfo.markDevirt(); 823 } 824 825 bool DevirtModule::tryUniformRetValOpt( 826 MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo, 827 WholeProgramDevirtResolution::ByArg *Res) { 828 // Uniform return value optimization. If all functions return the same 829 // constant, replace all calls with that constant. 830 uint64_t TheRetVal = TargetsForSlot[0].RetVal; 831 for (const VirtualCallTarget &Target : TargetsForSlot) 832 if (Target.RetVal != TheRetVal) 833 return false; 834 835 if (CSInfo.isExported()) { 836 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniformRetVal; 837 Res->Info = TheRetVal; 838 } 839 840 applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal); 841 if (RemarksEnabled) 842 for (auto &&Target : TargetsForSlot) 843 Target.WasDevirt = true; 844 return true; 845 } 846 847 std::string DevirtModule::getGlobalName(VTableSlot Slot, 848 ArrayRef<uint64_t> Args, 849 StringRef Name) { 850 std::string FullName = "__typeid_"; 851 raw_string_ostream OS(FullName); 852 OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset; 853 for (uint64_t Arg : Args) 854 OS << '_' << Arg; 855 OS << '_' << Name; 856 return OS.str(); 857 } 858 859 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 860 StringRef Name, Constant *C) { 861 GlobalAlias *GA = GlobalAlias::create(Int8Ty, 0, GlobalValue::ExternalLinkage, 862 getGlobalName(Slot, Args, Name), C, &M); 863 GA->setVisibility(GlobalValue::HiddenVisibility); 864 } 865 866 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 867 StringRef Name, unsigned AbsWidth) { 868 Constant *C = M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Ty); 869 auto *GV = dyn_cast<GlobalVariable>(C); 870 // We only need to set metadata if the global is newly created, in which 871 // case it would not have hidden visibility. 872 if (!GV || GV->getVisibility() == GlobalValue::HiddenVisibility) 873 return C; 874 875 GV->setVisibility(GlobalValue::HiddenVisibility); 876 auto SetAbsRange = [&](uint64_t Min, uint64_t Max) { 877 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min)); 878 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max)); 879 GV->setMetadata(LLVMContext::MD_absolute_symbol, 880 MDNode::get(M.getContext(), {MinC, MaxC})); 881 }; 882 if (AbsWidth == IntPtrTy->getBitWidth()) 883 SetAbsRange(~0ull, ~0ull); // Full set. 884 else if (AbsWidth) 885 SetAbsRange(0, 1ull << AbsWidth); 886 return GV; 887 } 888 889 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 890 bool IsOne, 891 Constant *UniqueMemberAddr) { 892 for (auto &&Call : CSInfo.CallSites) { 893 IRBuilder<> B(Call.CS.getInstruction()); 894 Value *Cmp = 895 B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE, 896 B.CreateBitCast(Call.VTable, Int8PtrTy), UniqueMemberAddr); 897 Cmp = B.CreateZExt(Cmp, Call.CS->getType()); 898 Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter, 899 Cmp); 900 } 901 CSInfo.markDevirt(); 902 } 903 904 bool DevirtModule::tryUniqueRetValOpt( 905 unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot, 906 CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res, 907 VTableSlot Slot, ArrayRef<uint64_t> Args) { 908 // IsOne controls whether we look for a 0 or a 1. 909 auto tryUniqueRetValOptFor = [&](bool IsOne) { 910 const TypeMemberInfo *UniqueMember = nullptr; 911 for (const VirtualCallTarget &Target : TargetsForSlot) { 912 if (Target.RetVal == (IsOne ? 1 : 0)) { 913 if (UniqueMember) 914 return false; 915 UniqueMember = Target.TM; 916 } 917 } 918 919 // We should have found a unique member or bailed out by now. We already 920 // checked for a uniform return value in tryUniformRetValOpt. 921 assert(UniqueMember); 922 923 Constant *UniqueMemberAddr = 924 ConstantExpr::getBitCast(UniqueMember->Bits->GV, Int8PtrTy); 925 UniqueMemberAddr = ConstantExpr::getGetElementPtr( 926 Int8Ty, UniqueMemberAddr, 927 ConstantInt::get(Int64Ty, UniqueMember->Offset)); 928 929 if (CSInfo.isExported()) { 930 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniqueRetVal; 931 Res->Info = IsOne; 932 933 exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr); 934 } 935 936 // Replace each call with the comparison. 937 applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne, 938 UniqueMemberAddr); 939 940 // Update devirtualization statistics for targets. 941 if (RemarksEnabled) 942 for (auto &&Target : TargetsForSlot) 943 Target.WasDevirt = true; 944 945 return true; 946 }; 947 948 if (BitWidth == 1) { 949 if (tryUniqueRetValOptFor(true)) 950 return true; 951 if (tryUniqueRetValOptFor(false)) 952 return true; 953 } 954 return false; 955 } 956 957 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 958 Constant *Byte, Constant *Bit) { 959 for (auto Call : CSInfo.CallSites) { 960 auto *RetType = cast<IntegerType>(Call.CS.getType()); 961 IRBuilder<> B(Call.CS.getInstruction()); 962 Value *Addr = 963 B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte); 964 if (RetType->getBitWidth() == 1) { 965 Value *Bits = B.CreateLoad(Addr); 966 Value *BitsAndBit = B.CreateAnd(Bits, Bit); 967 auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0)); 968 Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled, 969 OREGetter, IsBitSet); 970 } else { 971 Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo()); 972 Value *Val = B.CreateLoad(RetType, ValAddr); 973 Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled, 974 OREGetter, Val); 975 } 976 } 977 CSInfo.markDevirt(); 978 } 979 980 bool DevirtModule::tryVirtualConstProp( 981 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 982 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 983 // This only works if the function returns an integer. 984 auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType()); 985 if (!RetType) 986 return false; 987 unsigned BitWidth = RetType->getBitWidth(); 988 if (BitWidth > 64) 989 return false; 990 991 // Make sure that each function is defined, does not access memory, takes at 992 // least one argument, does not use its first argument (which we assume is 993 // 'this'), and has the same return type. 994 // 995 // Note that we test whether this copy of the function is readnone, rather 996 // than testing function attributes, which must hold for any copy of the 997 // function, even a less optimized version substituted at link time. This is 998 // sound because the virtual constant propagation optimizations effectively 999 // inline all implementations of the virtual function into each call site, 1000 // rather than using function attributes to perform local optimization. 1001 for (VirtualCallTarget &Target : TargetsForSlot) { 1002 if (Target.Fn->isDeclaration() || 1003 computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) != 1004 MAK_ReadNone || 1005 Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() || 1006 Target.Fn->getReturnType() != RetType) 1007 return false; 1008 } 1009 1010 for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) { 1011 if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first)) 1012 continue; 1013 1014 WholeProgramDevirtResolution::ByArg *ResByArg = nullptr; 1015 if (Res) 1016 ResByArg = &Res->ResByArg[CSByConstantArg.first]; 1017 1018 if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg)) 1019 continue; 1020 1021 if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second, 1022 ResByArg, Slot, CSByConstantArg.first)) 1023 continue; 1024 1025 // Find an allocation offset in bits in all vtables associated with the 1026 // type. 1027 uint64_t AllocBefore = 1028 findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth); 1029 uint64_t AllocAfter = 1030 findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth); 1031 1032 // Calculate the total amount of padding needed to store a value at both 1033 // ends of the object. 1034 uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0; 1035 for (auto &&Target : TargetsForSlot) { 1036 TotalPaddingBefore += std::max<int64_t>( 1037 (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0); 1038 TotalPaddingAfter += std::max<int64_t>( 1039 (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0); 1040 } 1041 1042 // If the amount of padding is too large, give up. 1043 // FIXME: do something smarter here. 1044 if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128) 1045 continue; 1046 1047 // Calculate the offset to the value as a (possibly negative) byte offset 1048 // and (if applicable) a bit offset, and store the values in the targets. 1049 int64_t OffsetByte; 1050 uint64_t OffsetBit; 1051 if (TotalPaddingBefore <= TotalPaddingAfter) 1052 setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte, 1053 OffsetBit); 1054 else 1055 setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte, 1056 OffsetBit); 1057 1058 if (RemarksEnabled) 1059 for (auto &&Target : TargetsForSlot) 1060 Target.WasDevirt = true; 1061 1062 Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte); 1063 Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit); 1064 1065 if (CSByConstantArg.second.isExported()) { 1066 ResByArg->TheKind = WholeProgramDevirtResolution::ByArg::VirtualConstProp; 1067 exportGlobal(Slot, CSByConstantArg.first, "byte", 1068 ConstantExpr::getIntToPtr(ByteConst, Int8PtrTy)); 1069 exportGlobal(Slot, CSByConstantArg.first, "bit", 1070 ConstantExpr::getIntToPtr(BitConst, Int8PtrTy)); 1071 } 1072 1073 // Rewrite each call to a load from OffsetByte/OffsetBit. 1074 applyVirtualConstProp(CSByConstantArg.second, 1075 TargetsForSlot[0].Fn->getName(), ByteConst, BitConst); 1076 } 1077 return true; 1078 } 1079 1080 void DevirtModule::rebuildGlobal(VTableBits &B) { 1081 if (B.Before.Bytes.empty() && B.After.Bytes.empty()) 1082 return; 1083 1084 // Align each byte array to pointer width. 1085 unsigned PointerSize = M.getDataLayout().getPointerSize(); 1086 B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), PointerSize)); 1087 B.After.Bytes.resize(alignTo(B.After.Bytes.size(), PointerSize)); 1088 1089 // Before was stored in reverse order; flip it now. 1090 for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I) 1091 std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]); 1092 1093 // Build an anonymous global containing the before bytes, followed by the 1094 // original initializer, followed by the after bytes. 1095 auto NewInit = ConstantStruct::getAnon( 1096 {ConstantDataArray::get(M.getContext(), B.Before.Bytes), 1097 B.GV->getInitializer(), 1098 ConstantDataArray::get(M.getContext(), B.After.Bytes)}); 1099 auto NewGV = 1100 new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(), 1101 GlobalVariable::PrivateLinkage, NewInit, "", B.GV); 1102 NewGV->setSection(B.GV->getSection()); 1103 NewGV->setComdat(B.GV->getComdat()); 1104 1105 // Copy the original vtable's metadata to the anonymous global, adjusting 1106 // offsets as required. 1107 NewGV->copyMetadata(B.GV, B.Before.Bytes.size()); 1108 1109 // Build an alias named after the original global, pointing at the second 1110 // element (the original initializer). 1111 auto Alias = GlobalAlias::create( 1112 B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "", 1113 ConstantExpr::getGetElementPtr( 1114 NewInit->getType(), NewGV, 1115 ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0), 1116 ConstantInt::get(Int32Ty, 1)}), 1117 &M); 1118 Alias->setVisibility(B.GV->getVisibility()); 1119 Alias->takeName(B.GV); 1120 1121 B.GV->replaceAllUsesWith(Alias); 1122 B.GV->eraseFromParent(); 1123 } 1124 1125 bool DevirtModule::areRemarksEnabled() { 1126 const auto &FL = M.getFunctionList(); 1127 if (FL.empty()) 1128 return false; 1129 const Function &Fn = FL.front(); 1130 1131 const auto &BBL = Fn.getBasicBlockList(); 1132 if (BBL.empty()) 1133 return false; 1134 auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front()); 1135 return DI.isEnabled(); 1136 } 1137 1138 void DevirtModule::scanTypeTestUsers(Function *TypeTestFunc, 1139 Function *AssumeFunc) { 1140 // Find all virtual calls via a virtual table pointer %p under an assumption 1141 // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p 1142 // points to a member of the type identifier %md. Group calls by (type ID, 1143 // offset) pair (effectively the identity of the virtual function) and store 1144 // to CallSlots. 1145 DenseSet<Value *> SeenPtrs; 1146 for (auto I = TypeTestFunc->use_begin(), E = TypeTestFunc->use_end(); 1147 I != E;) { 1148 auto CI = dyn_cast<CallInst>(I->getUser()); 1149 ++I; 1150 if (!CI) 1151 continue; 1152 1153 // Search for virtual calls based on %p and add them to DevirtCalls. 1154 SmallVector<DevirtCallSite, 1> DevirtCalls; 1155 SmallVector<CallInst *, 1> Assumes; 1156 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI); 1157 1158 // If we found any, add them to CallSlots. Only do this if we haven't seen 1159 // the vtable pointer before, as it may have been CSE'd with pointers from 1160 // other call sites, and we don't want to process call sites multiple times. 1161 if (!Assumes.empty()) { 1162 Metadata *TypeId = 1163 cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata(); 1164 Value *Ptr = CI->getArgOperand(0)->stripPointerCasts(); 1165 if (SeenPtrs.insert(Ptr).second) { 1166 for (DevirtCallSite Call : DevirtCalls) { 1167 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS, nullptr); 1168 } 1169 } 1170 } 1171 1172 // We no longer need the assumes or the type test. 1173 for (auto Assume : Assumes) 1174 Assume->eraseFromParent(); 1175 // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we 1176 // may use the vtable argument later. 1177 if (CI->use_empty()) 1178 CI->eraseFromParent(); 1179 } 1180 } 1181 1182 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) { 1183 Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test); 1184 1185 for (auto I = TypeCheckedLoadFunc->use_begin(), 1186 E = TypeCheckedLoadFunc->use_end(); 1187 I != E;) { 1188 auto CI = dyn_cast<CallInst>(I->getUser()); 1189 ++I; 1190 if (!CI) 1191 continue; 1192 1193 Value *Ptr = CI->getArgOperand(0); 1194 Value *Offset = CI->getArgOperand(1); 1195 Value *TypeIdValue = CI->getArgOperand(2); 1196 Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata(); 1197 1198 SmallVector<DevirtCallSite, 1> DevirtCalls; 1199 SmallVector<Instruction *, 1> LoadedPtrs; 1200 SmallVector<Instruction *, 1> Preds; 1201 bool HasNonCallUses = false; 1202 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds, 1203 HasNonCallUses, CI); 1204 1205 // Start by generating "pessimistic" code that explicitly loads the function 1206 // pointer from the vtable and performs the type check. If possible, we will 1207 // eliminate the load and the type check later. 1208 1209 // If possible, only generate the load at the point where it is used. 1210 // This helps avoid unnecessary spills. 1211 IRBuilder<> LoadB( 1212 (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI); 1213 Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset); 1214 Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy)); 1215 Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr); 1216 1217 for (Instruction *LoadedPtr : LoadedPtrs) { 1218 LoadedPtr->replaceAllUsesWith(LoadedValue); 1219 LoadedPtr->eraseFromParent(); 1220 } 1221 1222 // Likewise for the type test. 1223 IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI); 1224 CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue}); 1225 1226 for (Instruction *Pred : Preds) { 1227 Pred->replaceAllUsesWith(TypeTestCall); 1228 Pred->eraseFromParent(); 1229 } 1230 1231 // We have already erased any extractvalue instructions that refer to the 1232 // intrinsic call, but the intrinsic may have other non-extractvalue uses 1233 // (although this is unlikely). In that case, explicitly build a pair and 1234 // RAUW it. 1235 if (!CI->use_empty()) { 1236 Value *Pair = UndefValue::get(CI->getType()); 1237 IRBuilder<> B(CI); 1238 Pair = B.CreateInsertValue(Pair, LoadedValue, {0}); 1239 Pair = B.CreateInsertValue(Pair, TypeTestCall, {1}); 1240 CI->replaceAllUsesWith(Pair); 1241 } 1242 1243 // The number of unsafe uses is initially the number of uses. 1244 auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall]; 1245 NumUnsafeUses = DevirtCalls.size(); 1246 1247 // If the function pointer has a non-call user, we cannot eliminate the type 1248 // check, as one of those users may eventually call the pointer. Increment 1249 // the unsafe use count to make sure it cannot reach zero. 1250 if (HasNonCallUses) 1251 ++NumUnsafeUses; 1252 for (DevirtCallSite Call : DevirtCalls) { 1253 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS, 1254 &NumUnsafeUses); 1255 } 1256 1257 CI->eraseFromParent(); 1258 } 1259 } 1260 1261 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) { 1262 const TypeIdSummary *TidSummary = 1263 ImportSummary->getTypeIdSummary(cast<MDString>(Slot.TypeID)->getString()); 1264 if (!TidSummary) 1265 return; 1266 auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset); 1267 if (ResI == TidSummary->WPDRes.end()) 1268 return; 1269 const WholeProgramDevirtResolution &Res = ResI->second; 1270 1271 if (Res.TheKind == WholeProgramDevirtResolution::SingleImpl) { 1272 // The type of the function in the declaration is irrelevant because every 1273 // call site will cast it to the correct type. 1274 auto *SingleImpl = M.getOrInsertFunction( 1275 Res.SingleImplName, Type::getVoidTy(M.getContext())); 1276 1277 // This is the import phase so we should not be exporting anything. 1278 bool IsExported = false; 1279 applySingleImplDevirt(SlotInfo, SingleImpl, IsExported); 1280 assert(!IsExported); 1281 } 1282 1283 for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) { 1284 auto I = Res.ResByArg.find(CSByConstantArg.first); 1285 if (I == Res.ResByArg.end()) 1286 continue; 1287 auto &ResByArg = I->second; 1288 // FIXME: We should figure out what to do about the "function name" argument 1289 // to the apply* functions, as the function names are unavailable during the 1290 // importing phase. For now we just pass the empty string. This does not 1291 // impact correctness because the function names are just used for remarks. 1292 switch (ResByArg.TheKind) { 1293 case WholeProgramDevirtResolution::ByArg::UniformRetVal: 1294 applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info); 1295 break; 1296 case WholeProgramDevirtResolution::ByArg::UniqueRetVal: { 1297 Constant *UniqueMemberAddr = 1298 importGlobal(Slot, CSByConstantArg.first, "unique_member"); 1299 applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info, 1300 UniqueMemberAddr); 1301 break; 1302 } 1303 case WholeProgramDevirtResolution::ByArg::VirtualConstProp: { 1304 Constant *Byte = importGlobal(Slot, CSByConstantArg.first, "byte", 32); 1305 Byte = ConstantExpr::getPtrToInt(Byte, Int32Ty); 1306 Constant *Bit = importGlobal(Slot, CSByConstantArg.first, "bit", 8); 1307 Bit = ConstantExpr::getPtrToInt(Bit, Int8Ty); 1308 applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit); 1309 } 1310 default: 1311 break; 1312 } 1313 } 1314 } 1315 1316 void DevirtModule::removeRedundantTypeTests() { 1317 auto True = ConstantInt::getTrue(M.getContext()); 1318 for (auto &&U : NumUnsafeUsesForTypeTest) { 1319 if (U.second == 0) { 1320 U.first->replaceAllUsesWith(True); 1321 U.first->eraseFromParent(); 1322 } 1323 } 1324 } 1325 1326 bool DevirtModule::run() { 1327 Function *TypeTestFunc = 1328 M.getFunction(Intrinsic::getName(Intrinsic::type_test)); 1329 Function *TypeCheckedLoadFunc = 1330 M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load)); 1331 Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume)); 1332 1333 // Normally if there are no users of the devirtualization intrinsics in the 1334 // module, this pass has nothing to do. But if we are exporting, we also need 1335 // to handle any users that appear only in the function summaries. 1336 if (!ExportSummary && 1337 (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc || 1338 AssumeFunc->use_empty()) && 1339 (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty())) 1340 return false; 1341 1342 if (TypeTestFunc && AssumeFunc) 1343 scanTypeTestUsers(TypeTestFunc, AssumeFunc); 1344 1345 if (TypeCheckedLoadFunc) 1346 scanTypeCheckedLoadUsers(TypeCheckedLoadFunc); 1347 1348 if (ImportSummary) { 1349 for (auto &S : CallSlots) 1350 importResolution(S.first, S.second); 1351 1352 removeRedundantTypeTests(); 1353 1354 // The rest of the code is only necessary when exporting or during regular 1355 // LTO, so we are done. 1356 return true; 1357 } 1358 1359 // Rebuild type metadata into a map for easy lookup. 1360 std::vector<VTableBits> Bits; 1361 DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap; 1362 buildTypeIdentifierMap(Bits, TypeIdMap); 1363 if (TypeIdMap.empty()) 1364 return true; 1365 1366 // Collect information from summary about which calls to try to devirtualize. 1367 if (ExportSummary) { 1368 DenseMap<GlobalValue::GUID, TinyPtrVector<Metadata *>> MetadataByGUID; 1369 for (auto &P : TypeIdMap) { 1370 if (auto *TypeId = dyn_cast<MDString>(P.first)) 1371 MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back( 1372 TypeId); 1373 } 1374 1375 for (auto &P : *ExportSummary) { 1376 for (auto &S : P.second.SummaryList) { 1377 auto *FS = dyn_cast<FunctionSummary>(S.get()); 1378 if (!FS) 1379 continue; 1380 // FIXME: Only add live functions. 1381 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 1382 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 1383 CallSlots[{MD, VF.Offset}].CSInfo.SummaryHasTypeTestAssumeUsers = 1384 true; 1385 } 1386 } 1387 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 1388 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 1389 CallSlots[{MD, VF.Offset}] 1390 .CSInfo.SummaryTypeCheckedLoadUsers.push_back(FS); 1391 } 1392 } 1393 for (const FunctionSummary::ConstVCall &VC : 1394 FS->type_test_assume_const_vcalls()) { 1395 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 1396 CallSlots[{MD, VC.VFunc.Offset}] 1397 .ConstCSInfo[VC.Args] 1398 .SummaryHasTypeTestAssumeUsers = true; 1399 } 1400 } 1401 for (const FunctionSummary::ConstVCall &VC : 1402 FS->type_checked_load_const_vcalls()) { 1403 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 1404 CallSlots[{MD, VC.VFunc.Offset}] 1405 .ConstCSInfo[VC.Args] 1406 .SummaryTypeCheckedLoadUsers.push_back(FS); 1407 } 1408 } 1409 } 1410 } 1411 } 1412 1413 // For each (type, offset) pair: 1414 bool DidVirtualConstProp = false; 1415 std::map<std::string, Function*> DevirtTargets; 1416 for (auto &S : CallSlots) { 1417 // Search each of the members of the type identifier for the virtual 1418 // function implementation at offset S.first.ByteOffset, and add to 1419 // TargetsForSlot. 1420 std::vector<VirtualCallTarget> TargetsForSlot; 1421 if (tryFindVirtualCallTargets(TargetsForSlot, TypeIdMap[S.first.TypeID], 1422 S.first.ByteOffset)) { 1423 WholeProgramDevirtResolution *Res = nullptr; 1424 if (ExportSummary && isa<MDString>(S.first.TypeID)) 1425 Res = &ExportSummary 1426 ->getOrInsertTypeIdSummary( 1427 cast<MDString>(S.first.TypeID)->getString()) 1428 .WPDRes[S.first.ByteOffset]; 1429 1430 if (!trySingleImplDevirt(TargetsForSlot, S.second, Res) && 1431 tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first)) 1432 DidVirtualConstProp = true; 1433 1434 // Collect functions devirtualized at least for one call site for stats. 1435 if (RemarksEnabled) 1436 for (const auto &T : TargetsForSlot) 1437 if (T.WasDevirt) 1438 DevirtTargets[T.Fn->getName()] = T.Fn; 1439 } 1440 1441 // CFI-specific: if we are exporting and any llvm.type.checked.load 1442 // intrinsics were *not* devirtualized, we need to add the resulting 1443 // llvm.type.test intrinsics to the function summaries so that the 1444 // LowerTypeTests pass will export them. 1445 if (ExportSummary && isa<MDString>(S.first.TypeID)) { 1446 auto GUID = 1447 GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString()); 1448 for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers) 1449 FS->addTypeTest(GUID); 1450 for (auto &CCS : S.second.ConstCSInfo) 1451 for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers) 1452 FS->addTypeTest(GUID); 1453 } 1454 } 1455 1456 if (RemarksEnabled) { 1457 // Generate remarks for each devirtualized function. 1458 for (const auto &DT : DevirtTargets) { 1459 Function *F = DT.second; 1460 1461 // In the new pass manager, we can request the optimization 1462 // remark emitter pass on a per-function-basis, which the 1463 // OREGetter will do for us. 1464 // In the old pass manager, this is harder, so we just build 1465 // a optimization remark emitter on the fly, when we need it. 1466 std::unique_ptr<OptimizationRemarkEmitter> OwnedORE; 1467 OptimizationRemarkEmitter *ORE; 1468 if (OREGetter) 1469 ORE = &OREGetter(F); 1470 else { 1471 OwnedORE = make_unique<OptimizationRemarkEmitter>(F); 1472 ORE = OwnedORE.get(); 1473 } 1474 1475 using namespace ore; 1476 ORE->emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F) 1477 << "devirtualized " << NV("FunctionName", F->getName())); 1478 } 1479 } 1480 1481 removeRedundantTypeTests(); 1482 1483 // Rebuild each global we touched as part of virtual constant propagation to 1484 // include the before and after bytes. 1485 if (DidVirtualConstProp) 1486 for (VTableBits &B : Bits) 1487 rebuildGlobal(B); 1488 1489 return true; 1490 } 1491