1 //===- WholeProgramDevirt.cpp - Whole program virtual call optimization ---===// 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 implements whole program optimization of virtual calls in cases 10 // where we know (via !type metadata) that the list of callees is fixed. This 11 // includes the following: 12 // - Single implementation devirtualization: if a virtual call has a single 13 // possible callee, replace all calls with a direct call to that callee. 14 // - Virtual constant propagation: if the virtual function's return type is an 15 // integer <=64 bits and all possible callees are readnone, for each class and 16 // each list of constant arguments: evaluate the function, store the return 17 // value alongside the virtual table, and rewrite each virtual call as a load 18 // from the virtual table. 19 // - Uniform return value optimization: if the conditions for virtual constant 20 // propagation hold and each function returns the same constant value, replace 21 // each virtual call with that constant. 22 // - Unique return value optimization for i1 return values: if the conditions 23 // for virtual constant propagation hold and a single vtable's function 24 // returns 0, or a single vtable's function returns 1, replace each virtual 25 // call with a comparison of the vptr against that vtable's address. 26 // 27 // This pass is intended to be used during the regular and thin LTO pipelines: 28 // 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). 33 // 34 // During hybrid Regular/ThinLTO, the pass operates in two phases: 35 // - Export phase: this is run during the thin link over a single merged module 36 // that contains all vtables with !type metadata that participate in the link. 37 // The pass computes a resolution for each virtual call and stores it in the 38 // type identifier summary. 39 // - Import phase: this is run during the thin backends over the individual 40 // modules. The pass applies the resolutions previously computed during the 41 // import phase to each eligible virtual call. 42 // 43 // During ThinLTO, the pass operates in two phases: 44 // - Export phase: this is run during the thin link over the index which 45 // contains a summary of all vtables with !type metadata that participate in 46 // the link. It computes a resolution for each virtual call and stores it in 47 // the type identifier summary. Only single implementation devirtualization 48 // is supported. 49 // - Import phase: (same as with hybrid case above). 50 // 51 //===----------------------------------------------------------------------===// 52 53 #include "llvm/Transforms/IPO/WholeProgramDevirt.h" 54 #include "llvm/ADT/ArrayRef.h" 55 #include "llvm/ADT/DenseMap.h" 56 #include "llvm/ADT/DenseMapInfo.h" 57 #include "llvm/ADT/DenseSet.h" 58 #include "llvm/ADT/MapVector.h" 59 #include "llvm/ADT/SmallVector.h" 60 #include "llvm/ADT/Triple.h" 61 #include "llvm/ADT/iterator_range.h" 62 #include "llvm/Analysis/AssumptionCache.h" 63 #include "llvm/Analysis/BasicAliasAnalysis.h" 64 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 65 #include "llvm/Analysis/TypeMetadataUtils.h" 66 #include "llvm/Bitcode/BitcodeReader.h" 67 #include "llvm/Bitcode/BitcodeWriter.h" 68 #include "llvm/IR/Constants.h" 69 #include "llvm/IR/DataLayout.h" 70 #include "llvm/IR/DebugLoc.h" 71 #include "llvm/IR/DerivedTypes.h" 72 #include "llvm/IR/Dominators.h" 73 #include "llvm/IR/Function.h" 74 #include "llvm/IR/GlobalAlias.h" 75 #include "llvm/IR/GlobalVariable.h" 76 #include "llvm/IR/IRBuilder.h" 77 #include "llvm/IR/InstrTypes.h" 78 #include "llvm/IR/Instruction.h" 79 #include "llvm/IR/Instructions.h" 80 #include "llvm/IR/Intrinsics.h" 81 #include "llvm/IR/LLVMContext.h" 82 #include "llvm/IR/Metadata.h" 83 #include "llvm/IR/Module.h" 84 #include "llvm/IR/ModuleSummaryIndexYAML.h" 85 #include "llvm/InitializePasses.h" 86 #include "llvm/Pass.h" 87 #include "llvm/PassRegistry.h" 88 #include "llvm/Support/Casting.h" 89 #include "llvm/Support/CommandLine.h" 90 #include "llvm/Support/Errc.h" 91 #include "llvm/Support/Error.h" 92 #include "llvm/Support/FileSystem.h" 93 #include "llvm/Support/GlobPattern.h" 94 #include "llvm/Support/MathExtras.h" 95 #include "llvm/Transforms/IPO.h" 96 #include "llvm/Transforms/IPO/FunctionAttrs.h" 97 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 98 #include "llvm/Transforms/Utils/Evaluator.h" 99 #include <algorithm> 100 #include <cstddef> 101 #include <map> 102 #include <set> 103 #include <string> 104 105 using namespace llvm; 106 using namespace wholeprogramdevirt; 107 108 #define DEBUG_TYPE "wholeprogramdevirt" 109 110 static cl::opt<PassSummaryAction> ClSummaryAction( 111 "wholeprogramdevirt-summary-action", 112 cl::desc("What to do with the summary when running this pass"), 113 cl::values(clEnumValN(PassSummaryAction::None, "none", "Do nothing"), 114 clEnumValN(PassSummaryAction::Import, "import", 115 "Import typeid resolutions from summary and globals"), 116 clEnumValN(PassSummaryAction::Export, "export", 117 "Export typeid resolutions to summary and globals")), 118 cl::Hidden); 119 120 static cl::opt<std::string> ClReadSummary( 121 "wholeprogramdevirt-read-summary", 122 cl::desc( 123 "Read summary from given bitcode or YAML file before running pass"), 124 cl::Hidden); 125 126 static cl::opt<std::string> ClWriteSummary( 127 "wholeprogramdevirt-write-summary", 128 cl::desc("Write summary to given bitcode or YAML file after running pass. " 129 "Output file format is deduced from extension: *.bc means writing " 130 "bitcode, otherwise YAML"), 131 cl::Hidden); 132 133 static cl::opt<unsigned> 134 ClThreshold("wholeprogramdevirt-branch-funnel-threshold", cl::Hidden, 135 cl::init(10), cl::ZeroOrMore, 136 cl::desc("Maximum number of call targets per " 137 "call site to enable branch funnels")); 138 139 static cl::opt<bool> 140 PrintSummaryDevirt("wholeprogramdevirt-print-index-based", cl::Hidden, 141 cl::init(false), cl::ZeroOrMore, 142 cl::desc("Print index-based devirtualization messages")); 143 144 /// Provide a way to force enable whole program visibility in tests. 145 /// This is needed to support legacy tests that don't contain 146 /// !vcall_visibility metadata (the mere presense of type tests 147 /// previously implied hidden visibility). 148 static cl::opt<bool> 149 WholeProgramVisibility("whole-program-visibility", cl::init(false), 150 cl::Hidden, cl::ZeroOrMore, 151 cl::desc("Enable whole program visibility")); 152 153 /// Provide a way to force disable whole program for debugging or workarounds, 154 /// when enabled via the linker. 155 static cl::opt<bool> DisableWholeProgramVisibility( 156 "disable-whole-program-visibility", cl::init(false), cl::Hidden, 157 cl::ZeroOrMore, 158 cl::desc("Disable whole program visibility (overrides enabling options)")); 159 160 /// Provide way to prevent certain function from being devirtualized 161 static cl::list<std::string> 162 SkipFunctionNames("wholeprogramdevirt-skip", 163 cl::desc("Prevent function(s) from being devirtualized"), 164 cl::Hidden, cl::ZeroOrMore, cl::CommaSeparated); 165 166 /// Mechanism to add runtime checking of devirtualization decisions, trapping on 167 /// any that are not correct. Useful for debugging undefined behavior leading to 168 /// failures with WPD. 169 static cl::opt<bool> 170 CheckDevirt("wholeprogramdevirt-check", cl::init(false), cl::Hidden, 171 cl::ZeroOrMore, 172 cl::desc("Add code to trap on incorrect devirtualizations")); 173 174 namespace { 175 struct PatternList { 176 std::vector<GlobPattern> Patterns; 177 template <class T> void init(const T &StringList) { 178 for (const auto &S : StringList) 179 if (Expected<GlobPattern> Pat = GlobPattern::create(S)) 180 Patterns.push_back(std::move(*Pat)); 181 } 182 bool match(StringRef S) { 183 for (const GlobPattern &P : Patterns) 184 if (P.match(S)) 185 return true; 186 return false; 187 } 188 }; 189 } // namespace 190 191 // Find the minimum offset that we may store a value of size Size bits at. If 192 // IsAfter is set, look for an offset before the object, otherwise look for an 193 // offset after the object. 194 uint64_t 195 wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets, 196 bool IsAfter, uint64_t Size) { 197 // Find a minimum offset taking into account only vtable sizes. 198 uint64_t MinByte = 0; 199 for (const VirtualCallTarget &Target : Targets) { 200 if (IsAfter) 201 MinByte = std::max(MinByte, Target.minAfterBytes()); 202 else 203 MinByte = std::max(MinByte, Target.minBeforeBytes()); 204 } 205 206 // Build a vector of arrays of bytes covering, for each target, a slice of the 207 // used region (see AccumBitVector::BytesUsed in 208 // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively, 209 // this aligns the used regions to start at MinByte. 210 // 211 // In this example, A, B and C are vtables, # is a byte already allocated for 212 // a virtual function pointer, AAAA... (etc.) are the used regions for the 213 // vtables and Offset(X) is the value computed for the Offset variable below 214 // for X. 215 // 216 // Offset(A) 217 // | | 218 // |MinByte 219 // A: ################AAAAAAAA|AAAAAAAA 220 // B: ########BBBBBBBBBBBBBBBB|BBBB 221 // C: ########################|CCCCCCCCCCCCCCCC 222 // | Offset(B) | 223 // 224 // This code produces the slices of A, B and C that appear after the divider 225 // at MinByte. 226 std::vector<ArrayRef<uint8_t>> Used; 227 for (const VirtualCallTarget &Target : Targets) { 228 ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed 229 : Target.TM->Bits->Before.BytesUsed; 230 uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes() 231 : MinByte - Target.minBeforeBytes(); 232 233 // Disregard used regions that are smaller than Offset. These are 234 // effectively all-free regions that do not need to be checked. 235 if (VTUsed.size() > Offset) 236 Used.push_back(VTUsed.slice(Offset)); 237 } 238 239 if (Size == 1) { 240 // Find a free bit in each member of Used. 241 for (unsigned I = 0;; ++I) { 242 uint8_t BitsUsed = 0; 243 for (auto &&B : Used) 244 if (I < B.size()) 245 BitsUsed |= B[I]; 246 if (BitsUsed != 0xff) 247 return (MinByte + I) * 8 + 248 countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined); 249 } 250 } else { 251 // Find a free (Size/8) byte region in each member of Used. 252 // FIXME: see if alignment helps. 253 for (unsigned I = 0;; ++I) { 254 for (auto &&B : Used) { 255 unsigned Byte = 0; 256 while ((I + Byte) < B.size() && Byte < (Size / 8)) { 257 if (B[I + Byte]) 258 goto NextI; 259 ++Byte; 260 } 261 } 262 return (MinByte + I) * 8; 263 NextI:; 264 } 265 } 266 } 267 268 void wholeprogramdevirt::setBeforeReturnValues( 269 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore, 270 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 271 if (BitWidth == 1) 272 OffsetByte = -(AllocBefore / 8 + 1); 273 else 274 OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8); 275 OffsetBit = AllocBefore % 8; 276 277 for (VirtualCallTarget &Target : Targets) { 278 if (BitWidth == 1) 279 Target.setBeforeBit(AllocBefore); 280 else 281 Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8); 282 } 283 } 284 285 void wholeprogramdevirt::setAfterReturnValues( 286 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter, 287 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 288 if (BitWidth == 1) 289 OffsetByte = AllocAfter / 8; 290 else 291 OffsetByte = (AllocAfter + 7) / 8; 292 OffsetBit = AllocAfter % 8; 293 294 for (VirtualCallTarget &Target : Targets) { 295 if (BitWidth == 1) 296 Target.setAfterBit(AllocAfter); 297 else 298 Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8); 299 } 300 } 301 302 VirtualCallTarget::VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM) 303 : Fn(Fn), TM(TM), 304 IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {} 305 306 namespace { 307 308 // A slot in a set of virtual tables. The TypeID identifies the set of virtual 309 // tables, and the ByteOffset is the offset in bytes from the address point to 310 // the virtual function pointer. 311 struct VTableSlot { 312 Metadata *TypeID; 313 uint64_t ByteOffset; 314 }; 315 316 } // end anonymous namespace 317 318 namespace llvm { 319 320 template <> struct DenseMapInfo<VTableSlot> { 321 static VTableSlot getEmptyKey() { 322 return {DenseMapInfo<Metadata *>::getEmptyKey(), 323 DenseMapInfo<uint64_t>::getEmptyKey()}; 324 } 325 static VTableSlot getTombstoneKey() { 326 return {DenseMapInfo<Metadata *>::getTombstoneKey(), 327 DenseMapInfo<uint64_t>::getTombstoneKey()}; 328 } 329 static unsigned getHashValue(const VTableSlot &I) { 330 return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^ 331 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 332 } 333 static bool isEqual(const VTableSlot &LHS, 334 const VTableSlot &RHS) { 335 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 336 } 337 }; 338 339 template <> struct DenseMapInfo<VTableSlotSummary> { 340 static VTableSlotSummary getEmptyKey() { 341 return {DenseMapInfo<StringRef>::getEmptyKey(), 342 DenseMapInfo<uint64_t>::getEmptyKey()}; 343 } 344 static VTableSlotSummary getTombstoneKey() { 345 return {DenseMapInfo<StringRef>::getTombstoneKey(), 346 DenseMapInfo<uint64_t>::getTombstoneKey()}; 347 } 348 static unsigned getHashValue(const VTableSlotSummary &I) { 349 return DenseMapInfo<StringRef>::getHashValue(I.TypeID) ^ 350 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 351 } 352 static bool isEqual(const VTableSlotSummary &LHS, 353 const VTableSlotSummary &RHS) { 354 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 355 } 356 }; 357 358 } // end namespace llvm 359 360 namespace { 361 362 // Returns true if the function must be unreachable based on ValueInfo. 363 // 364 // In particular, identifies a function as unreachable in the following 365 // conditions 366 // 1) All summaries are live. 367 // 2) All function summaries indicate it's unreachable 368 bool mustBeUnreachableFunction(ValueInfo TheFnVI) { 369 if ((!TheFnVI) || TheFnVI.getSummaryList().empty()) { 370 // Returns false if ValueInfo is absent, or the summary list is empty 371 // (e.g., function declarations). 372 return false; 373 } 374 375 for (auto &Summary : TheFnVI.getSummaryList()) { 376 // Conservatively returns false if any non-live functions are seen. 377 // In general either all summaries should be live or all should be dead. 378 if (!Summary->isLive()) 379 return false; 380 if (auto *FS = dyn_cast<FunctionSummary>(Summary.get())) { 381 if (!FS->fflags().MustBeUnreachable) 382 return false; 383 } 384 // Do nothing if a non-function has the same GUID (which is rare). 385 // This is correct since non-function summaries are not relevant. 386 } 387 // All function summaries are live and all of them agree that the function is 388 // unreachble. 389 return true; 390 } 391 392 // A virtual call site. VTable is the loaded virtual table pointer, and CS is 393 // the indirect virtual call. 394 struct VirtualCallSite { 395 Value *VTable = nullptr; 396 CallBase &CB; 397 398 // If non-null, this field points to the associated unsafe use count stored in 399 // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description 400 // of that field for details. 401 unsigned *NumUnsafeUses = nullptr; 402 403 void 404 emitRemark(const StringRef OptName, const StringRef TargetName, 405 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) { 406 Function *F = CB.getCaller(); 407 DebugLoc DLoc = CB.getDebugLoc(); 408 BasicBlock *Block = CB.getParent(); 409 410 using namespace ore; 411 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, OptName, DLoc, Block) 412 << NV("Optimization", OptName) 413 << ": devirtualized a call to " 414 << NV("FunctionName", TargetName)); 415 } 416 417 void replaceAndErase( 418 const StringRef OptName, const StringRef TargetName, bool RemarksEnabled, 419 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 420 Value *New) { 421 if (RemarksEnabled) 422 emitRemark(OptName, TargetName, OREGetter); 423 CB.replaceAllUsesWith(New); 424 if (auto *II = dyn_cast<InvokeInst>(&CB)) { 425 BranchInst::Create(II->getNormalDest(), &CB); 426 II->getUnwindDest()->removePredecessor(II->getParent()); 427 } 428 CB.eraseFromParent(); 429 // This use is no longer unsafe. 430 if (NumUnsafeUses) 431 --*NumUnsafeUses; 432 } 433 }; 434 435 // Call site information collected for a specific VTableSlot and possibly a list 436 // of constant integer arguments. The grouping by arguments is handled by the 437 // VTableSlotInfo class. 438 struct CallSiteInfo { 439 /// The set of call sites for this slot. Used during regular LTO and the 440 /// import phase of ThinLTO (as well as the export phase of ThinLTO for any 441 /// call sites that appear in the merged module itself); in each of these 442 /// cases we are directly operating on the call sites at the IR level. 443 std::vector<VirtualCallSite> CallSites; 444 445 /// Whether all call sites represented by this CallSiteInfo, including those 446 /// in summaries, have been devirtualized. This starts off as true because a 447 /// default constructed CallSiteInfo represents no call sites. 448 bool AllCallSitesDevirted = true; 449 450 // These fields are used during the export phase of ThinLTO and reflect 451 // information collected from function summaries. 452 453 /// Whether any function summary contains an llvm.assume(llvm.type.test) for 454 /// this slot. 455 bool SummaryHasTypeTestAssumeUsers = false; 456 457 /// CFI-specific: a vector containing the list of function summaries that use 458 /// the llvm.type.checked.load intrinsic and therefore will require 459 /// resolutions for llvm.type.test in order to implement CFI checks if 460 /// devirtualization was unsuccessful. If devirtualization was successful, the 461 /// pass will clear this vector by calling markDevirt(). If at the end of the 462 /// pass the vector is non-empty, we will need to add a use of llvm.type.test 463 /// to each of the function summaries in the vector. 464 std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers; 465 std::vector<FunctionSummary *> SummaryTypeTestAssumeUsers; 466 467 bool isExported() const { 468 return SummaryHasTypeTestAssumeUsers || 469 !SummaryTypeCheckedLoadUsers.empty(); 470 } 471 472 void addSummaryTypeCheckedLoadUser(FunctionSummary *FS) { 473 SummaryTypeCheckedLoadUsers.push_back(FS); 474 AllCallSitesDevirted = false; 475 } 476 477 void addSummaryTypeTestAssumeUser(FunctionSummary *FS) { 478 SummaryTypeTestAssumeUsers.push_back(FS); 479 SummaryHasTypeTestAssumeUsers = true; 480 AllCallSitesDevirted = false; 481 } 482 483 void markDevirt() { 484 AllCallSitesDevirted = true; 485 486 // As explained in the comment for SummaryTypeCheckedLoadUsers. 487 SummaryTypeCheckedLoadUsers.clear(); 488 } 489 }; 490 491 // Call site information collected for a specific VTableSlot. 492 struct VTableSlotInfo { 493 // The set of call sites which do not have all constant integer arguments 494 // (excluding "this"). 495 CallSiteInfo CSInfo; 496 497 // The set of call sites with all constant integer arguments (excluding 498 // "this"), grouped by argument list. 499 std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo; 500 501 void addCallSite(Value *VTable, CallBase &CB, unsigned *NumUnsafeUses); 502 503 private: 504 CallSiteInfo &findCallSiteInfo(CallBase &CB); 505 }; 506 507 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallBase &CB) { 508 std::vector<uint64_t> Args; 509 auto *CBType = dyn_cast<IntegerType>(CB.getType()); 510 if (!CBType || CBType->getBitWidth() > 64 || CB.arg_empty()) 511 return CSInfo; 512 for (auto &&Arg : drop_begin(CB.args())) { 513 auto *CI = dyn_cast<ConstantInt>(Arg); 514 if (!CI || CI->getBitWidth() > 64) 515 return CSInfo; 516 Args.push_back(CI->getZExtValue()); 517 } 518 return ConstCSInfo[Args]; 519 } 520 521 void VTableSlotInfo::addCallSite(Value *VTable, CallBase &CB, 522 unsigned *NumUnsafeUses) { 523 auto &CSI = findCallSiteInfo(CB); 524 CSI.AllCallSitesDevirted = false; 525 CSI.CallSites.push_back({VTable, CB, NumUnsafeUses}); 526 } 527 528 struct DevirtModule { 529 Module &M; 530 function_ref<AAResults &(Function &)> AARGetter; 531 function_ref<DominatorTree &(Function &)> LookupDomTree; 532 533 ModuleSummaryIndex *ExportSummary; 534 const ModuleSummaryIndex *ImportSummary; 535 536 IntegerType *Int8Ty; 537 PointerType *Int8PtrTy; 538 IntegerType *Int32Ty; 539 IntegerType *Int64Ty; 540 IntegerType *IntPtrTy; 541 /// Sizeless array type, used for imported vtables. This provides a signal 542 /// to analyzers that these imports may alias, as they do for example 543 /// when multiple unique return values occur in the same vtable. 544 ArrayType *Int8Arr0Ty; 545 546 bool RemarksEnabled; 547 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter; 548 549 MapVector<VTableSlot, VTableSlotInfo> CallSlots; 550 551 // Calls that have already been optimized. We may add a call to multiple 552 // VTableSlotInfos if vtable loads are coalesced and need to make sure not to 553 // optimize a call more than once. 554 SmallPtrSet<CallBase *, 8> OptimizedCalls; 555 556 // This map keeps track of the number of "unsafe" uses of a loaded function 557 // pointer. The key is the associated llvm.type.test intrinsic call generated 558 // by this pass. An unsafe use is one that calls the loaded function pointer 559 // directly. Every time we eliminate an unsafe use (for example, by 560 // devirtualizing it or by applying virtual constant propagation), we 561 // decrement the value stored in this map. If a value reaches zero, we can 562 // eliminate the type check by RAUWing the associated llvm.type.test call with 563 // true. 564 std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest; 565 PatternList FunctionsToSkip; 566 567 DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter, 568 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 569 function_ref<DominatorTree &(Function &)> LookupDomTree, 570 ModuleSummaryIndex *ExportSummary, 571 const ModuleSummaryIndex *ImportSummary) 572 : M(M), AARGetter(AARGetter), LookupDomTree(LookupDomTree), 573 ExportSummary(ExportSummary), ImportSummary(ImportSummary), 574 Int8Ty(Type::getInt8Ty(M.getContext())), 575 Int8PtrTy(Type::getInt8PtrTy(M.getContext())), 576 Int32Ty(Type::getInt32Ty(M.getContext())), 577 Int64Ty(Type::getInt64Ty(M.getContext())), 578 IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)), 579 Int8Arr0Ty(ArrayType::get(Type::getInt8Ty(M.getContext()), 0)), 580 RemarksEnabled(areRemarksEnabled()), OREGetter(OREGetter) { 581 assert(!(ExportSummary && ImportSummary)); 582 FunctionsToSkip.init(SkipFunctionNames); 583 } 584 585 bool areRemarksEnabled(); 586 587 void 588 scanTypeTestUsers(Function *TypeTestFunc, 589 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap); 590 void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc); 591 592 void buildTypeIdentifierMap( 593 std::vector<VTableBits> &Bits, 594 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap); 595 596 bool 597 tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot, 598 const std::set<TypeMemberInfo> &TypeMemberInfos, 599 uint64_t ByteOffset, 600 ModuleSummaryIndex *ExportSummary); 601 602 void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn, 603 bool &IsExported); 604 bool trySingleImplDevirt(ModuleSummaryIndex *ExportSummary, 605 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 606 VTableSlotInfo &SlotInfo, 607 WholeProgramDevirtResolution *Res); 608 609 void applyICallBranchFunnel(VTableSlotInfo &SlotInfo, Constant *JT, 610 bool &IsExported); 611 void tryICallBranchFunnel(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 612 VTableSlotInfo &SlotInfo, 613 WholeProgramDevirtResolution *Res, VTableSlot Slot); 614 615 bool tryEvaluateFunctionsWithArgs( 616 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 617 ArrayRef<uint64_t> Args); 618 619 void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 620 uint64_t TheRetVal); 621 bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 622 CallSiteInfo &CSInfo, 623 WholeProgramDevirtResolution::ByArg *Res); 624 625 // Returns the global symbol name that is used to export information about the 626 // given vtable slot and list of arguments. 627 std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args, 628 StringRef Name); 629 630 bool shouldExportConstantsAsAbsoluteSymbols(); 631 632 // This function is called during the export phase to create a symbol 633 // definition containing information about the given vtable slot and list of 634 // arguments. 635 void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 636 Constant *C); 637 void exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 638 uint32_t Const, uint32_t &Storage); 639 640 // This function is called during the import phase to create a reference to 641 // the symbol definition created during the export phase. 642 Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 643 StringRef Name); 644 Constant *importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 645 StringRef Name, IntegerType *IntTy, 646 uint32_t Storage); 647 648 Constant *getMemberAddr(const TypeMemberInfo *M); 649 650 void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne, 651 Constant *UniqueMemberAddr); 652 bool tryUniqueRetValOpt(unsigned BitWidth, 653 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 654 CallSiteInfo &CSInfo, 655 WholeProgramDevirtResolution::ByArg *Res, 656 VTableSlot Slot, ArrayRef<uint64_t> Args); 657 658 void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 659 Constant *Byte, Constant *Bit); 660 bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 661 VTableSlotInfo &SlotInfo, 662 WholeProgramDevirtResolution *Res, VTableSlot Slot); 663 664 void rebuildGlobal(VTableBits &B); 665 666 // Apply the summary resolution for Slot to all virtual calls in SlotInfo. 667 void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo); 668 669 // If we were able to eliminate all unsafe uses for a type checked load, 670 // eliminate the associated type tests by replacing them with true. 671 void removeRedundantTypeTests(); 672 673 bool run(); 674 675 // Look up the corresponding ValueInfo entry of `TheFn` in `ExportSummary`. 676 // 677 // Caller guarantees that `ExportSummary` is not nullptr. 678 static ValueInfo lookUpFunctionValueInfo(Function *TheFn, 679 ModuleSummaryIndex *ExportSummary); 680 681 // Returns true if the function definition must be unreachable. 682 // 683 // Note if this helper function returns true, `F` is guaranteed 684 // to be unreachable; if it returns false, `F` might still 685 // be unreachable but not covered by this helper function. 686 // 687 // Implementation-wise, if function definition is present, IR is analyzed; if 688 // not, look up function flags from ExportSummary as a fallback. 689 static bool mustBeUnreachableFunction(Function *const F, 690 ModuleSummaryIndex *ExportSummary); 691 692 // Lower the module using the action and summary passed as command line 693 // arguments. For testing purposes only. 694 static bool 695 runForTesting(Module &M, function_ref<AAResults &(Function &)> AARGetter, 696 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 697 function_ref<DominatorTree &(Function &)> LookupDomTree); 698 }; 699 700 struct DevirtIndex { 701 ModuleSummaryIndex &ExportSummary; 702 // The set in which to record GUIDs exported from their module by 703 // devirtualization, used by client to ensure they are not internalized. 704 std::set<GlobalValue::GUID> &ExportedGUIDs; 705 // A map in which to record the information necessary to locate the WPD 706 // resolution for local targets in case they are exported by cross module 707 // importing. 708 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap; 709 710 MapVector<VTableSlotSummary, VTableSlotInfo> CallSlots; 711 712 PatternList FunctionsToSkip; 713 714 DevirtIndex( 715 ModuleSummaryIndex &ExportSummary, 716 std::set<GlobalValue::GUID> &ExportedGUIDs, 717 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) 718 : ExportSummary(ExportSummary), ExportedGUIDs(ExportedGUIDs), 719 LocalWPDTargetsMap(LocalWPDTargetsMap) { 720 FunctionsToSkip.init(SkipFunctionNames); 721 } 722 723 bool tryFindVirtualCallTargets(std::vector<ValueInfo> &TargetsForSlot, 724 const TypeIdCompatibleVtableInfo TIdInfo, 725 uint64_t ByteOffset); 726 727 bool trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 728 VTableSlotSummary &SlotSummary, 729 VTableSlotInfo &SlotInfo, 730 WholeProgramDevirtResolution *Res, 731 std::set<ValueInfo> &DevirtTargets); 732 733 void run(); 734 }; 735 736 struct WholeProgramDevirt : public ModulePass { 737 static char ID; 738 739 bool UseCommandLine = false; 740 741 ModuleSummaryIndex *ExportSummary = nullptr; 742 const ModuleSummaryIndex *ImportSummary = nullptr; 743 744 WholeProgramDevirt() : ModulePass(ID), UseCommandLine(true) { 745 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry()); 746 } 747 748 WholeProgramDevirt(ModuleSummaryIndex *ExportSummary, 749 const ModuleSummaryIndex *ImportSummary) 750 : ModulePass(ID), ExportSummary(ExportSummary), 751 ImportSummary(ImportSummary) { 752 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry()); 753 } 754 755 bool runOnModule(Module &M) override { 756 if (skipModule(M)) 757 return false; 758 759 // In the new pass manager, we can request the optimization 760 // remark emitter pass on a per-function-basis, which the 761 // OREGetter will do for us. 762 // In the old pass manager, this is harder, so we just build 763 // an optimization remark emitter on the fly, when we need it. 764 std::unique_ptr<OptimizationRemarkEmitter> ORE; 765 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { 766 ORE = std::make_unique<OptimizationRemarkEmitter>(F); 767 return *ORE; 768 }; 769 770 auto LookupDomTree = [this](Function &F) -> DominatorTree & { 771 return this->getAnalysis<DominatorTreeWrapperPass>(F).getDomTree(); 772 }; 773 774 if (UseCommandLine) 775 return DevirtModule::runForTesting(M, LegacyAARGetter(*this), OREGetter, 776 LookupDomTree); 777 778 return DevirtModule(M, LegacyAARGetter(*this), OREGetter, LookupDomTree, 779 ExportSummary, ImportSummary) 780 .run(); 781 } 782 783 void getAnalysisUsage(AnalysisUsage &AU) const override { 784 AU.addRequired<AssumptionCacheTracker>(); 785 AU.addRequired<TargetLibraryInfoWrapperPass>(); 786 AU.addRequired<DominatorTreeWrapperPass>(); 787 } 788 }; 789 790 } // end anonymous namespace 791 792 INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt", 793 "Whole program devirtualization", false, false) 794 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 795 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 796 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 797 INITIALIZE_PASS_END(WholeProgramDevirt, "wholeprogramdevirt", 798 "Whole program devirtualization", false, false) 799 char WholeProgramDevirt::ID = 0; 800 801 ModulePass * 802 llvm::createWholeProgramDevirtPass(ModuleSummaryIndex *ExportSummary, 803 const ModuleSummaryIndex *ImportSummary) { 804 return new WholeProgramDevirt(ExportSummary, ImportSummary); 805 } 806 807 PreservedAnalyses WholeProgramDevirtPass::run(Module &M, 808 ModuleAnalysisManager &AM) { 809 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 810 auto AARGetter = [&](Function &F) -> AAResults & { 811 return FAM.getResult<AAManager>(F); 812 }; 813 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { 814 return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F); 815 }; 816 auto LookupDomTree = [&FAM](Function &F) -> DominatorTree & { 817 return FAM.getResult<DominatorTreeAnalysis>(F); 818 }; 819 if (UseCommandLine) { 820 if (DevirtModule::runForTesting(M, AARGetter, OREGetter, LookupDomTree)) 821 return PreservedAnalyses::all(); 822 return PreservedAnalyses::none(); 823 } 824 if (!DevirtModule(M, AARGetter, OREGetter, LookupDomTree, ExportSummary, 825 ImportSummary) 826 .run()) 827 return PreservedAnalyses::all(); 828 return PreservedAnalyses::none(); 829 } 830 831 // Enable whole program visibility if enabled by client (e.g. linker) or 832 // internal option, and not force disabled. 833 static bool hasWholeProgramVisibility(bool WholeProgramVisibilityEnabledInLTO) { 834 return (WholeProgramVisibilityEnabledInLTO || WholeProgramVisibility) && 835 !DisableWholeProgramVisibility; 836 } 837 838 namespace llvm { 839 840 /// If whole program visibility asserted, then upgrade all public vcall 841 /// visibility metadata on vtable definitions to linkage unit visibility in 842 /// Module IR (for regular or hybrid LTO). 843 void updateVCallVisibilityInModule( 844 Module &M, bool WholeProgramVisibilityEnabledInLTO, 845 const DenseSet<GlobalValue::GUID> &DynamicExportSymbols) { 846 if (!hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO)) 847 return; 848 for (GlobalVariable &GV : M.globals()) 849 // Add linkage unit visibility to any variable with type metadata, which are 850 // the vtable definitions. We won't have an existing vcall_visibility 851 // metadata on vtable definitions with public visibility. 852 if (GV.hasMetadata(LLVMContext::MD_type) && 853 GV.getVCallVisibility() == GlobalObject::VCallVisibilityPublic && 854 // Don't upgrade the visibility for symbols exported to the dynamic 855 // linker, as we have no information on their eventual use. 856 !DynamicExportSymbols.count(GV.getGUID())) 857 GV.setVCallVisibilityMetadata(GlobalObject::VCallVisibilityLinkageUnit); 858 } 859 860 /// If whole program visibility asserted, then upgrade all public vcall 861 /// visibility metadata on vtable definition summaries to linkage unit 862 /// visibility in Module summary index (for ThinLTO). 863 void updateVCallVisibilityInIndex( 864 ModuleSummaryIndex &Index, bool WholeProgramVisibilityEnabledInLTO, 865 const DenseSet<GlobalValue::GUID> &DynamicExportSymbols) { 866 if (!hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO)) 867 return; 868 for (auto &P : Index) { 869 for (auto &S : P.second.SummaryList) { 870 auto *GVar = dyn_cast<GlobalVarSummary>(S.get()); 871 if (!GVar || 872 GVar->getVCallVisibility() != GlobalObject::VCallVisibilityPublic || 873 // Don't upgrade the visibility for symbols exported to the dynamic 874 // linker, as we have no information on their eventual use. 875 DynamicExportSymbols.count(P.first)) 876 continue; 877 GVar->setVCallVisibility(GlobalObject::VCallVisibilityLinkageUnit); 878 } 879 } 880 } 881 882 void runWholeProgramDevirtOnIndex( 883 ModuleSummaryIndex &Summary, std::set<GlobalValue::GUID> &ExportedGUIDs, 884 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 885 DevirtIndex(Summary, ExportedGUIDs, LocalWPDTargetsMap).run(); 886 } 887 888 void updateIndexWPDForExports( 889 ModuleSummaryIndex &Summary, 890 function_ref<bool(StringRef, ValueInfo)> isExported, 891 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 892 for (auto &T : LocalWPDTargetsMap) { 893 auto &VI = T.first; 894 // This was enforced earlier during trySingleImplDevirt. 895 assert(VI.getSummaryList().size() == 1 && 896 "Devirt of local target has more than one copy"); 897 auto &S = VI.getSummaryList()[0]; 898 if (!isExported(S->modulePath(), VI)) 899 continue; 900 901 // It's been exported by a cross module import. 902 for (auto &SlotSummary : T.second) { 903 auto *TIdSum = Summary.getTypeIdSummary(SlotSummary.TypeID); 904 assert(TIdSum); 905 auto WPDRes = TIdSum->WPDRes.find(SlotSummary.ByteOffset); 906 assert(WPDRes != TIdSum->WPDRes.end()); 907 WPDRes->second.SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 908 WPDRes->second.SingleImplName, 909 Summary.getModuleHash(S->modulePath())); 910 } 911 } 912 } 913 914 } // end namespace llvm 915 916 static Error checkCombinedSummaryForTesting(ModuleSummaryIndex *Summary) { 917 // Check that summary index contains regular LTO module when performing 918 // export to prevent occasional use of index from pure ThinLTO compilation 919 // (-fno-split-lto-module). This kind of summary index is passed to 920 // DevirtIndex::run, not to DevirtModule::run used by opt/runForTesting. 921 const auto &ModPaths = Summary->modulePaths(); 922 if (ClSummaryAction != PassSummaryAction::Import && 923 ModPaths.find(ModuleSummaryIndex::getRegularLTOModuleName()) == 924 ModPaths.end()) 925 return createStringError( 926 errc::invalid_argument, 927 "combined summary should contain Regular LTO module"); 928 return ErrorSuccess(); 929 } 930 931 bool DevirtModule::runForTesting( 932 Module &M, function_ref<AAResults &(Function &)> AARGetter, 933 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 934 function_ref<DominatorTree &(Function &)> LookupDomTree) { 935 std::unique_ptr<ModuleSummaryIndex> Summary = 936 std::make_unique<ModuleSummaryIndex>(/*HaveGVs=*/false); 937 938 // Handle the command-line summary arguments. This code is for testing 939 // purposes only, so we handle errors directly. 940 if (!ClReadSummary.empty()) { 941 ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary + 942 ": "); 943 auto ReadSummaryFile = 944 ExitOnErr(errorOrToExpected(MemoryBuffer::getFile(ClReadSummary))); 945 if (Expected<std::unique_ptr<ModuleSummaryIndex>> SummaryOrErr = 946 getModuleSummaryIndex(*ReadSummaryFile)) { 947 Summary = std::move(*SummaryOrErr); 948 ExitOnErr(checkCombinedSummaryForTesting(Summary.get())); 949 } else { 950 // Try YAML if we've failed with bitcode. 951 consumeError(SummaryOrErr.takeError()); 952 yaml::Input In(ReadSummaryFile->getBuffer()); 953 In >> *Summary; 954 ExitOnErr(errorCodeToError(In.error())); 955 } 956 } 957 958 bool Changed = 959 DevirtModule(M, AARGetter, OREGetter, LookupDomTree, 960 ClSummaryAction == PassSummaryAction::Export ? Summary.get() 961 : nullptr, 962 ClSummaryAction == PassSummaryAction::Import ? Summary.get() 963 : nullptr) 964 .run(); 965 966 if (!ClWriteSummary.empty()) { 967 ExitOnError ExitOnErr( 968 "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": "); 969 std::error_code EC; 970 if (StringRef(ClWriteSummary).endswith(".bc")) { 971 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_None); 972 ExitOnErr(errorCodeToError(EC)); 973 writeIndexToFile(*Summary, OS); 974 } else { 975 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_TextWithCRLF); 976 ExitOnErr(errorCodeToError(EC)); 977 yaml::Output Out(OS); 978 Out << *Summary; 979 } 980 } 981 982 return Changed; 983 } 984 985 void DevirtModule::buildTypeIdentifierMap( 986 std::vector<VTableBits> &Bits, 987 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 988 DenseMap<GlobalVariable *, VTableBits *> GVToBits; 989 Bits.reserve(M.getGlobalList().size()); 990 SmallVector<MDNode *, 2> Types; 991 for (GlobalVariable &GV : M.globals()) { 992 Types.clear(); 993 GV.getMetadata(LLVMContext::MD_type, Types); 994 if (GV.isDeclaration() || Types.empty()) 995 continue; 996 997 VTableBits *&BitsPtr = GVToBits[&GV]; 998 if (!BitsPtr) { 999 Bits.emplace_back(); 1000 Bits.back().GV = &GV; 1001 Bits.back().ObjectSize = 1002 M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType()); 1003 BitsPtr = &Bits.back(); 1004 } 1005 1006 for (MDNode *Type : Types) { 1007 auto TypeID = Type->getOperand(1).get(); 1008 1009 uint64_t Offset = 1010 cast<ConstantInt>( 1011 cast<ConstantAsMetadata>(Type->getOperand(0))->getValue()) 1012 ->getZExtValue(); 1013 1014 TypeIdMap[TypeID].insert({BitsPtr, Offset}); 1015 } 1016 } 1017 } 1018 1019 bool DevirtModule::tryFindVirtualCallTargets( 1020 std::vector<VirtualCallTarget> &TargetsForSlot, 1021 const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset, 1022 ModuleSummaryIndex *ExportSummary) { 1023 for (const TypeMemberInfo &TM : TypeMemberInfos) { 1024 if (!TM.Bits->GV->isConstant()) 1025 return false; 1026 1027 // We cannot perform whole program devirtualization analysis on a vtable 1028 // with public LTO visibility. 1029 if (TM.Bits->GV->getVCallVisibility() == 1030 GlobalObject::VCallVisibilityPublic) 1031 return false; 1032 1033 Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(), 1034 TM.Offset + ByteOffset, M); 1035 if (!Ptr) 1036 return false; 1037 1038 auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts()); 1039 if (!Fn) 1040 return false; 1041 1042 if (FunctionsToSkip.match(Fn->getName())) 1043 return false; 1044 1045 // We can disregard __cxa_pure_virtual as a possible call target, as 1046 // calls to pure virtuals are UB. 1047 if (Fn->getName() == "__cxa_pure_virtual") 1048 continue; 1049 1050 // We can disregard unreachable functions as possible call targets, as 1051 // unreachable functions shouldn't be called. 1052 if (mustBeUnreachableFunction(Fn, ExportSummary)) 1053 continue; 1054 1055 TargetsForSlot.push_back({Fn, &TM}); 1056 } 1057 1058 // Give up if we couldn't find any targets. 1059 return !TargetsForSlot.empty(); 1060 } 1061 1062 bool DevirtIndex::tryFindVirtualCallTargets( 1063 std::vector<ValueInfo> &TargetsForSlot, const TypeIdCompatibleVtableInfo TIdInfo, 1064 uint64_t ByteOffset) { 1065 for (const TypeIdOffsetVtableInfo &P : TIdInfo) { 1066 // Find a representative copy of the vtable initializer. 1067 // We can have multiple available_externally, linkonce_odr and weak_odr 1068 // vtable initializers. We can also have multiple external vtable 1069 // initializers in the case of comdats, which we cannot check here. 1070 // The linker should give an error in this case. 1071 // 1072 // Also, handle the case of same-named local Vtables with the same path 1073 // and therefore the same GUID. This can happen if there isn't enough 1074 // distinguishing path when compiling the source file. In that case we 1075 // conservatively return false early. 1076 const GlobalVarSummary *VS = nullptr; 1077 bool LocalFound = false; 1078 for (auto &S : P.VTableVI.getSummaryList()) { 1079 if (GlobalValue::isLocalLinkage(S->linkage())) { 1080 if (LocalFound) 1081 return false; 1082 LocalFound = true; 1083 } 1084 auto *CurVS = cast<GlobalVarSummary>(S->getBaseObject()); 1085 if (!CurVS->vTableFuncs().empty() || 1086 // Previously clang did not attach the necessary type metadata to 1087 // available_externally vtables, in which case there would not 1088 // be any vtable functions listed in the summary and we need 1089 // to treat this case conservatively (in case the bitcode is old). 1090 // However, we will also not have any vtable functions in the 1091 // case of a pure virtual base class. In that case we do want 1092 // to set VS to avoid treating it conservatively. 1093 !GlobalValue::isAvailableExternallyLinkage(S->linkage())) { 1094 VS = CurVS; 1095 // We cannot perform whole program devirtualization analysis on a vtable 1096 // with public LTO visibility. 1097 if (VS->getVCallVisibility() == GlobalObject::VCallVisibilityPublic) 1098 return false; 1099 } 1100 } 1101 // There will be no VS if all copies are available_externally having no 1102 // type metadata. In that case we can't safely perform WPD. 1103 if (!VS) 1104 return false; 1105 if (!VS->isLive()) 1106 continue; 1107 for (auto VTP : VS->vTableFuncs()) { 1108 if (VTP.VTableOffset != P.AddressPointOffset + ByteOffset) 1109 continue; 1110 1111 if (mustBeUnreachableFunction(VTP.FuncVI)) 1112 continue; 1113 1114 TargetsForSlot.push_back(VTP.FuncVI); 1115 } 1116 } 1117 1118 // Give up if we couldn't find any targets. 1119 return !TargetsForSlot.empty(); 1120 } 1121 1122 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo, 1123 Constant *TheFn, bool &IsExported) { 1124 // Don't devirtualize function if we're told to skip it 1125 // in -wholeprogramdevirt-skip. 1126 if (FunctionsToSkip.match(TheFn->stripPointerCasts()->getName())) 1127 return; 1128 auto Apply = [&](CallSiteInfo &CSInfo) { 1129 for (auto &&VCallSite : CSInfo.CallSites) { 1130 if (!OptimizedCalls.insert(&VCallSite.CB).second) 1131 continue; 1132 1133 if (RemarksEnabled) 1134 VCallSite.emitRemark("single-impl", 1135 TheFn->stripPointerCasts()->getName(), OREGetter); 1136 auto &CB = VCallSite.CB; 1137 assert(!CB.getCalledFunction() && "devirtualizing direct call?"); 1138 IRBuilder<> Builder(&CB); 1139 Value *Callee = 1140 Builder.CreateBitCast(TheFn, CB.getCalledOperand()->getType()); 1141 1142 // If checking is enabled, add support to compare the virtual function 1143 // pointer to the devirtualized target. In case of a mismatch, perform a 1144 // debug trap. 1145 if (CheckDevirt) { 1146 auto *Cond = Builder.CreateICmpNE(CB.getCalledOperand(), Callee); 1147 Instruction *ThenTerm = 1148 SplitBlockAndInsertIfThen(Cond, &CB, /*Unreachable=*/false); 1149 Builder.SetInsertPoint(ThenTerm); 1150 Function *TrapFn = Intrinsic::getDeclaration(&M, Intrinsic::debugtrap); 1151 auto *CallTrap = Builder.CreateCall(TrapFn); 1152 CallTrap->setDebugLoc(CB.getDebugLoc()); 1153 } 1154 1155 // Devirtualize. 1156 CB.setCalledOperand(Callee); 1157 1158 // This use is no longer unsafe. 1159 if (VCallSite.NumUnsafeUses) 1160 --*VCallSite.NumUnsafeUses; 1161 } 1162 if (CSInfo.isExported()) 1163 IsExported = true; 1164 CSInfo.markDevirt(); 1165 }; 1166 Apply(SlotInfo.CSInfo); 1167 for (auto &P : SlotInfo.ConstCSInfo) 1168 Apply(P.second); 1169 } 1170 1171 static bool AddCalls(VTableSlotInfo &SlotInfo, const ValueInfo &Callee) { 1172 // We can't add calls if we haven't seen a definition 1173 if (Callee.getSummaryList().empty()) 1174 return false; 1175 1176 // Insert calls into the summary index so that the devirtualized targets 1177 // are eligible for import. 1178 // FIXME: Annotate type tests with hotness. For now, mark these as hot 1179 // to better ensure we have the opportunity to inline them. 1180 bool IsExported = false; 1181 auto &S = Callee.getSummaryList()[0]; 1182 CalleeInfo CI(CalleeInfo::HotnessType::Hot, /* RelBF = */ 0); 1183 auto AddCalls = [&](CallSiteInfo &CSInfo) { 1184 for (auto *FS : CSInfo.SummaryTypeCheckedLoadUsers) { 1185 FS->addCall({Callee, CI}); 1186 IsExported |= S->modulePath() != FS->modulePath(); 1187 } 1188 for (auto *FS : CSInfo.SummaryTypeTestAssumeUsers) { 1189 FS->addCall({Callee, CI}); 1190 IsExported |= S->modulePath() != FS->modulePath(); 1191 } 1192 }; 1193 AddCalls(SlotInfo.CSInfo); 1194 for (auto &P : SlotInfo.ConstCSInfo) 1195 AddCalls(P.second); 1196 return IsExported; 1197 } 1198 1199 bool DevirtModule::trySingleImplDevirt( 1200 ModuleSummaryIndex *ExportSummary, 1201 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1202 WholeProgramDevirtResolution *Res) { 1203 // See if the program contains a single implementation of this virtual 1204 // function. 1205 Function *TheFn = TargetsForSlot[0].Fn; 1206 for (auto &&Target : TargetsForSlot) 1207 if (TheFn != Target.Fn) 1208 return false; 1209 1210 // If so, update each call site to call that implementation directly. 1211 if (RemarksEnabled) 1212 TargetsForSlot[0].WasDevirt = true; 1213 1214 bool IsExported = false; 1215 applySingleImplDevirt(SlotInfo, TheFn, IsExported); 1216 if (!IsExported) 1217 return false; 1218 1219 // If the only implementation has local linkage, we must promote to external 1220 // to make it visible to thin LTO objects. We can only get here during the 1221 // ThinLTO export phase. 1222 if (TheFn->hasLocalLinkage()) { 1223 std::string NewName = (TheFn->getName() + ".llvm.merged").str(); 1224 1225 // Since we are renaming the function, any comdats with the same name must 1226 // also be renamed. This is required when targeting COFF, as the comdat name 1227 // must match one of the names of the symbols in the comdat. 1228 if (Comdat *C = TheFn->getComdat()) { 1229 if (C->getName() == TheFn->getName()) { 1230 Comdat *NewC = M.getOrInsertComdat(NewName); 1231 NewC->setSelectionKind(C->getSelectionKind()); 1232 for (GlobalObject &GO : M.global_objects()) 1233 if (GO.getComdat() == C) 1234 GO.setComdat(NewC); 1235 } 1236 } 1237 1238 TheFn->setLinkage(GlobalValue::ExternalLinkage); 1239 TheFn->setVisibility(GlobalValue::HiddenVisibility); 1240 TheFn->setName(NewName); 1241 } 1242 if (ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFn->getGUID())) 1243 // Any needed promotion of 'TheFn' has already been done during 1244 // LTO unit split, so we can ignore return value of AddCalls. 1245 AddCalls(SlotInfo, TheFnVI); 1246 1247 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1248 Res->SingleImplName = std::string(TheFn->getName()); 1249 1250 return true; 1251 } 1252 1253 bool DevirtIndex::trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 1254 VTableSlotSummary &SlotSummary, 1255 VTableSlotInfo &SlotInfo, 1256 WholeProgramDevirtResolution *Res, 1257 std::set<ValueInfo> &DevirtTargets) { 1258 // See if the program contains a single implementation of this virtual 1259 // function. 1260 auto TheFn = TargetsForSlot[0]; 1261 for (auto &&Target : TargetsForSlot) 1262 if (TheFn != Target) 1263 return false; 1264 1265 // Don't devirtualize if we don't have target definition. 1266 auto Size = TheFn.getSummaryList().size(); 1267 if (!Size) 1268 return false; 1269 1270 // Don't devirtualize function if we're told to skip it 1271 // in -wholeprogramdevirt-skip. 1272 if (FunctionsToSkip.match(TheFn.name())) 1273 return false; 1274 1275 // If the summary list contains multiple summaries where at least one is 1276 // a local, give up, as we won't know which (possibly promoted) name to use. 1277 for (auto &S : TheFn.getSummaryList()) 1278 if (GlobalValue::isLocalLinkage(S->linkage()) && Size > 1) 1279 return false; 1280 1281 // Collect functions devirtualized at least for one call site for stats. 1282 if (PrintSummaryDevirt) 1283 DevirtTargets.insert(TheFn); 1284 1285 auto &S = TheFn.getSummaryList()[0]; 1286 bool IsExported = AddCalls(SlotInfo, TheFn); 1287 if (IsExported) 1288 ExportedGUIDs.insert(TheFn.getGUID()); 1289 1290 // Record in summary for use in devirtualization during the ThinLTO import 1291 // step. 1292 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1293 if (GlobalValue::isLocalLinkage(S->linkage())) { 1294 if (IsExported) 1295 // If target is a local function and we are exporting it by 1296 // devirtualizing a call in another module, we need to record the 1297 // promoted name. 1298 Res->SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 1299 TheFn.name(), ExportSummary.getModuleHash(S->modulePath())); 1300 else { 1301 LocalWPDTargetsMap[TheFn].push_back(SlotSummary); 1302 Res->SingleImplName = std::string(TheFn.name()); 1303 } 1304 } else 1305 Res->SingleImplName = std::string(TheFn.name()); 1306 1307 // Name will be empty if this thin link driven off of serialized combined 1308 // index (e.g. llvm-lto). However, WPD is not supported/invoked for the 1309 // legacy LTO API anyway. 1310 assert(!Res->SingleImplName.empty()); 1311 1312 return true; 1313 } 1314 1315 void DevirtModule::tryICallBranchFunnel( 1316 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1317 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1318 Triple T(M.getTargetTriple()); 1319 if (T.getArch() != Triple::x86_64) 1320 return; 1321 1322 if (TargetsForSlot.size() > ClThreshold) 1323 return; 1324 1325 bool HasNonDevirt = !SlotInfo.CSInfo.AllCallSitesDevirted; 1326 if (!HasNonDevirt) 1327 for (auto &P : SlotInfo.ConstCSInfo) 1328 if (!P.second.AllCallSitesDevirted) { 1329 HasNonDevirt = true; 1330 break; 1331 } 1332 1333 if (!HasNonDevirt) 1334 return; 1335 1336 FunctionType *FT = 1337 FunctionType::get(Type::getVoidTy(M.getContext()), {Int8PtrTy}, true); 1338 Function *JT; 1339 if (isa<MDString>(Slot.TypeID)) { 1340 JT = Function::Create(FT, Function::ExternalLinkage, 1341 M.getDataLayout().getProgramAddressSpace(), 1342 getGlobalName(Slot, {}, "branch_funnel"), &M); 1343 JT->setVisibility(GlobalValue::HiddenVisibility); 1344 } else { 1345 JT = Function::Create(FT, Function::InternalLinkage, 1346 M.getDataLayout().getProgramAddressSpace(), 1347 "branch_funnel", &M); 1348 } 1349 JT->addParamAttr(0, Attribute::Nest); 1350 1351 std::vector<Value *> JTArgs; 1352 JTArgs.push_back(JT->arg_begin()); 1353 for (auto &T : TargetsForSlot) { 1354 JTArgs.push_back(getMemberAddr(T.TM)); 1355 JTArgs.push_back(T.Fn); 1356 } 1357 1358 BasicBlock *BB = BasicBlock::Create(M.getContext(), "", JT, nullptr); 1359 Function *Intr = 1360 Intrinsic::getDeclaration(&M, llvm::Intrinsic::icall_branch_funnel, {}); 1361 1362 auto *CI = CallInst::Create(Intr, JTArgs, "", BB); 1363 CI->setTailCallKind(CallInst::TCK_MustTail); 1364 ReturnInst::Create(M.getContext(), nullptr, BB); 1365 1366 bool IsExported = false; 1367 applyICallBranchFunnel(SlotInfo, JT, IsExported); 1368 if (IsExported) 1369 Res->TheKind = WholeProgramDevirtResolution::BranchFunnel; 1370 } 1371 1372 void DevirtModule::applyICallBranchFunnel(VTableSlotInfo &SlotInfo, 1373 Constant *JT, bool &IsExported) { 1374 auto Apply = [&](CallSiteInfo &CSInfo) { 1375 if (CSInfo.isExported()) 1376 IsExported = true; 1377 if (CSInfo.AllCallSitesDevirted) 1378 return; 1379 for (auto &&VCallSite : CSInfo.CallSites) { 1380 CallBase &CB = VCallSite.CB; 1381 1382 // Jump tables are only profitable if the retpoline mitigation is enabled. 1383 Attribute FSAttr = CB.getCaller()->getFnAttribute("target-features"); 1384 if (!FSAttr.isValid() || 1385 !FSAttr.getValueAsString().contains("+retpoline")) 1386 continue; 1387 1388 if (RemarksEnabled) 1389 VCallSite.emitRemark("branch-funnel", 1390 JT->stripPointerCasts()->getName(), OREGetter); 1391 1392 // Pass the address of the vtable in the nest register, which is r10 on 1393 // x86_64. 1394 std::vector<Type *> NewArgs; 1395 NewArgs.push_back(Int8PtrTy); 1396 append_range(NewArgs, CB.getFunctionType()->params()); 1397 FunctionType *NewFT = 1398 FunctionType::get(CB.getFunctionType()->getReturnType(), NewArgs, 1399 CB.getFunctionType()->isVarArg()); 1400 PointerType *NewFTPtr = PointerType::getUnqual(NewFT); 1401 1402 IRBuilder<> IRB(&CB); 1403 std::vector<Value *> Args; 1404 Args.push_back(IRB.CreateBitCast(VCallSite.VTable, Int8PtrTy)); 1405 llvm::append_range(Args, CB.args()); 1406 1407 CallBase *NewCS = nullptr; 1408 if (isa<CallInst>(CB)) 1409 NewCS = IRB.CreateCall(NewFT, IRB.CreateBitCast(JT, NewFTPtr), Args); 1410 else 1411 NewCS = IRB.CreateInvoke(NewFT, IRB.CreateBitCast(JT, NewFTPtr), 1412 cast<InvokeInst>(CB).getNormalDest(), 1413 cast<InvokeInst>(CB).getUnwindDest(), Args); 1414 NewCS->setCallingConv(CB.getCallingConv()); 1415 1416 AttributeList Attrs = CB.getAttributes(); 1417 std::vector<AttributeSet> NewArgAttrs; 1418 NewArgAttrs.push_back(AttributeSet::get( 1419 M.getContext(), ArrayRef<Attribute>{Attribute::get( 1420 M.getContext(), Attribute::Nest)})); 1421 for (unsigned I = 0; I + 2 < Attrs.getNumAttrSets(); ++I) 1422 NewArgAttrs.push_back(Attrs.getParamAttrs(I)); 1423 NewCS->setAttributes( 1424 AttributeList::get(M.getContext(), Attrs.getFnAttrs(), 1425 Attrs.getRetAttrs(), NewArgAttrs)); 1426 1427 CB.replaceAllUsesWith(NewCS); 1428 CB.eraseFromParent(); 1429 1430 // This use is no longer unsafe. 1431 if (VCallSite.NumUnsafeUses) 1432 --*VCallSite.NumUnsafeUses; 1433 } 1434 // Don't mark as devirtualized because there may be callers compiled without 1435 // retpoline mitigation, which would mean that they are lowered to 1436 // llvm.type.test and therefore require an llvm.type.test resolution for the 1437 // type identifier. 1438 }; 1439 Apply(SlotInfo.CSInfo); 1440 for (auto &P : SlotInfo.ConstCSInfo) 1441 Apply(P.second); 1442 } 1443 1444 bool DevirtModule::tryEvaluateFunctionsWithArgs( 1445 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1446 ArrayRef<uint64_t> Args) { 1447 // Evaluate each function and store the result in each target's RetVal 1448 // field. 1449 for (VirtualCallTarget &Target : TargetsForSlot) { 1450 if (Target.Fn->arg_size() != Args.size() + 1) 1451 return false; 1452 1453 Evaluator Eval(M.getDataLayout(), nullptr); 1454 SmallVector<Constant *, 2> EvalArgs; 1455 EvalArgs.push_back( 1456 Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0))); 1457 for (unsigned I = 0; I != Args.size(); ++I) { 1458 auto *ArgTy = dyn_cast<IntegerType>( 1459 Target.Fn->getFunctionType()->getParamType(I + 1)); 1460 if (!ArgTy) 1461 return false; 1462 EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I])); 1463 } 1464 1465 Constant *RetVal; 1466 if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) || 1467 !isa<ConstantInt>(RetVal)) 1468 return false; 1469 Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue(); 1470 } 1471 return true; 1472 } 1473 1474 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1475 uint64_t TheRetVal) { 1476 for (auto Call : CSInfo.CallSites) { 1477 if (!OptimizedCalls.insert(&Call.CB).second) 1478 continue; 1479 Call.replaceAndErase( 1480 "uniform-ret-val", FnName, RemarksEnabled, OREGetter, 1481 ConstantInt::get(cast<IntegerType>(Call.CB.getType()), TheRetVal)); 1482 } 1483 CSInfo.markDevirt(); 1484 } 1485 1486 bool DevirtModule::tryUniformRetValOpt( 1487 MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo, 1488 WholeProgramDevirtResolution::ByArg *Res) { 1489 // Uniform return value optimization. If all functions return the same 1490 // constant, replace all calls with that constant. 1491 uint64_t TheRetVal = TargetsForSlot[0].RetVal; 1492 for (const VirtualCallTarget &Target : TargetsForSlot) 1493 if (Target.RetVal != TheRetVal) 1494 return false; 1495 1496 if (CSInfo.isExported()) { 1497 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniformRetVal; 1498 Res->Info = TheRetVal; 1499 } 1500 1501 applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal); 1502 if (RemarksEnabled) 1503 for (auto &&Target : TargetsForSlot) 1504 Target.WasDevirt = true; 1505 return true; 1506 } 1507 1508 std::string DevirtModule::getGlobalName(VTableSlot Slot, 1509 ArrayRef<uint64_t> Args, 1510 StringRef Name) { 1511 std::string FullName = "__typeid_"; 1512 raw_string_ostream OS(FullName); 1513 OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset; 1514 for (uint64_t Arg : Args) 1515 OS << '_' << Arg; 1516 OS << '_' << Name; 1517 return OS.str(); 1518 } 1519 1520 bool DevirtModule::shouldExportConstantsAsAbsoluteSymbols() { 1521 Triple T(M.getTargetTriple()); 1522 return T.isX86() && T.getObjectFormat() == Triple::ELF; 1523 } 1524 1525 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1526 StringRef Name, Constant *C) { 1527 GlobalAlias *GA = GlobalAlias::create(Int8Ty, 0, GlobalValue::ExternalLinkage, 1528 getGlobalName(Slot, Args, Name), C, &M); 1529 GA->setVisibility(GlobalValue::HiddenVisibility); 1530 } 1531 1532 void DevirtModule::exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1533 StringRef Name, uint32_t Const, 1534 uint32_t &Storage) { 1535 if (shouldExportConstantsAsAbsoluteSymbols()) { 1536 exportGlobal( 1537 Slot, Args, Name, 1538 ConstantExpr::getIntToPtr(ConstantInt::get(Int32Ty, Const), Int8PtrTy)); 1539 return; 1540 } 1541 1542 Storage = Const; 1543 } 1544 1545 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1546 StringRef Name) { 1547 Constant *C = 1548 M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Arr0Ty); 1549 auto *GV = dyn_cast<GlobalVariable>(C); 1550 if (GV) 1551 GV->setVisibility(GlobalValue::HiddenVisibility); 1552 return C; 1553 } 1554 1555 Constant *DevirtModule::importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1556 StringRef Name, IntegerType *IntTy, 1557 uint32_t Storage) { 1558 if (!shouldExportConstantsAsAbsoluteSymbols()) 1559 return ConstantInt::get(IntTy, Storage); 1560 1561 Constant *C = importGlobal(Slot, Args, Name); 1562 auto *GV = cast<GlobalVariable>(C->stripPointerCasts()); 1563 C = ConstantExpr::getPtrToInt(C, IntTy); 1564 1565 // We only need to set metadata if the global is newly created, in which 1566 // case it would not have hidden visibility. 1567 if (GV->hasMetadata(LLVMContext::MD_absolute_symbol)) 1568 return C; 1569 1570 auto SetAbsRange = [&](uint64_t Min, uint64_t Max) { 1571 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min)); 1572 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max)); 1573 GV->setMetadata(LLVMContext::MD_absolute_symbol, 1574 MDNode::get(M.getContext(), {MinC, MaxC})); 1575 }; 1576 unsigned AbsWidth = IntTy->getBitWidth(); 1577 if (AbsWidth == IntPtrTy->getBitWidth()) 1578 SetAbsRange(~0ull, ~0ull); // Full set. 1579 else 1580 SetAbsRange(0, 1ull << AbsWidth); 1581 return C; 1582 } 1583 1584 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1585 bool IsOne, 1586 Constant *UniqueMemberAddr) { 1587 for (auto &&Call : CSInfo.CallSites) { 1588 if (!OptimizedCalls.insert(&Call.CB).second) 1589 continue; 1590 IRBuilder<> B(&Call.CB); 1591 Value *Cmp = 1592 B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE, Call.VTable, 1593 B.CreateBitCast(UniqueMemberAddr, Call.VTable->getType())); 1594 Cmp = B.CreateZExt(Cmp, Call.CB.getType()); 1595 Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter, 1596 Cmp); 1597 } 1598 CSInfo.markDevirt(); 1599 } 1600 1601 Constant *DevirtModule::getMemberAddr(const TypeMemberInfo *M) { 1602 Constant *C = ConstantExpr::getBitCast(M->Bits->GV, Int8PtrTy); 1603 return ConstantExpr::getGetElementPtr(Int8Ty, C, 1604 ConstantInt::get(Int64Ty, M->Offset)); 1605 } 1606 1607 bool DevirtModule::tryUniqueRetValOpt( 1608 unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1609 CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res, 1610 VTableSlot Slot, ArrayRef<uint64_t> Args) { 1611 // IsOne controls whether we look for a 0 or a 1. 1612 auto tryUniqueRetValOptFor = [&](bool IsOne) { 1613 const TypeMemberInfo *UniqueMember = nullptr; 1614 for (const VirtualCallTarget &Target : TargetsForSlot) { 1615 if (Target.RetVal == (IsOne ? 1 : 0)) { 1616 if (UniqueMember) 1617 return false; 1618 UniqueMember = Target.TM; 1619 } 1620 } 1621 1622 // We should have found a unique member or bailed out by now. We already 1623 // checked for a uniform return value in tryUniformRetValOpt. 1624 assert(UniqueMember); 1625 1626 Constant *UniqueMemberAddr = getMemberAddr(UniqueMember); 1627 if (CSInfo.isExported()) { 1628 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniqueRetVal; 1629 Res->Info = IsOne; 1630 1631 exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr); 1632 } 1633 1634 // Replace each call with the comparison. 1635 applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne, 1636 UniqueMemberAddr); 1637 1638 // Update devirtualization statistics for targets. 1639 if (RemarksEnabled) 1640 for (auto &&Target : TargetsForSlot) 1641 Target.WasDevirt = true; 1642 1643 return true; 1644 }; 1645 1646 if (BitWidth == 1) { 1647 if (tryUniqueRetValOptFor(true)) 1648 return true; 1649 if (tryUniqueRetValOptFor(false)) 1650 return true; 1651 } 1652 return false; 1653 } 1654 1655 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 1656 Constant *Byte, Constant *Bit) { 1657 for (auto Call : CSInfo.CallSites) { 1658 if (!OptimizedCalls.insert(&Call.CB).second) 1659 continue; 1660 auto *RetType = cast<IntegerType>(Call.CB.getType()); 1661 IRBuilder<> B(&Call.CB); 1662 Value *Addr = 1663 B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte); 1664 if (RetType->getBitWidth() == 1) { 1665 Value *Bits = B.CreateLoad(Int8Ty, Addr); 1666 Value *BitsAndBit = B.CreateAnd(Bits, Bit); 1667 auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0)); 1668 Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled, 1669 OREGetter, IsBitSet); 1670 } else { 1671 Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo()); 1672 Value *Val = B.CreateLoad(RetType, ValAddr); 1673 Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled, 1674 OREGetter, Val); 1675 } 1676 } 1677 CSInfo.markDevirt(); 1678 } 1679 1680 bool DevirtModule::tryVirtualConstProp( 1681 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1682 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1683 // This only works if the function returns an integer. 1684 auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType()); 1685 if (!RetType) 1686 return false; 1687 unsigned BitWidth = RetType->getBitWidth(); 1688 if (BitWidth > 64) 1689 return false; 1690 1691 // Make sure that each function is defined, does not access memory, takes at 1692 // least one argument, does not use its first argument (which we assume is 1693 // 'this'), and has the same return type. 1694 // 1695 // Note that we test whether this copy of the function is readnone, rather 1696 // than testing function attributes, which must hold for any copy of the 1697 // function, even a less optimized version substituted at link time. This is 1698 // sound because the virtual constant propagation optimizations effectively 1699 // inline all implementations of the virtual function into each call site, 1700 // rather than using function attributes to perform local optimization. 1701 for (VirtualCallTarget &Target : TargetsForSlot) { 1702 if (Target.Fn->isDeclaration() || 1703 computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) != 1704 MAK_ReadNone || 1705 Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() || 1706 Target.Fn->getReturnType() != RetType) 1707 return false; 1708 } 1709 1710 for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) { 1711 if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first)) 1712 continue; 1713 1714 WholeProgramDevirtResolution::ByArg *ResByArg = nullptr; 1715 if (Res) 1716 ResByArg = &Res->ResByArg[CSByConstantArg.first]; 1717 1718 if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg)) 1719 continue; 1720 1721 if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second, 1722 ResByArg, Slot, CSByConstantArg.first)) 1723 continue; 1724 1725 // Find an allocation offset in bits in all vtables associated with the 1726 // type. 1727 uint64_t AllocBefore = 1728 findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth); 1729 uint64_t AllocAfter = 1730 findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth); 1731 1732 // Calculate the total amount of padding needed to store a value at both 1733 // ends of the object. 1734 uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0; 1735 for (auto &&Target : TargetsForSlot) { 1736 TotalPaddingBefore += std::max<int64_t>( 1737 (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0); 1738 TotalPaddingAfter += std::max<int64_t>( 1739 (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0); 1740 } 1741 1742 // If the amount of padding is too large, give up. 1743 // FIXME: do something smarter here. 1744 if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128) 1745 continue; 1746 1747 // Calculate the offset to the value as a (possibly negative) byte offset 1748 // and (if applicable) a bit offset, and store the values in the targets. 1749 int64_t OffsetByte; 1750 uint64_t OffsetBit; 1751 if (TotalPaddingBefore <= TotalPaddingAfter) 1752 setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte, 1753 OffsetBit); 1754 else 1755 setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte, 1756 OffsetBit); 1757 1758 if (RemarksEnabled) 1759 for (auto &&Target : TargetsForSlot) 1760 Target.WasDevirt = true; 1761 1762 1763 if (CSByConstantArg.second.isExported()) { 1764 ResByArg->TheKind = WholeProgramDevirtResolution::ByArg::VirtualConstProp; 1765 exportConstant(Slot, CSByConstantArg.first, "byte", OffsetByte, 1766 ResByArg->Byte); 1767 exportConstant(Slot, CSByConstantArg.first, "bit", 1ULL << OffsetBit, 1768 ResByArg->Bit); 1769 } 1770 1771 // Rewrite each call to a load from OffsetByte/OffsetBit. 1772 Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte); 1773 Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit); 1774 applyVirtualConstProp(CSByConstantArg.second, 1775 TargetsForSlot[0].Fn->getName(), ByteConst, BitConst); 1776 } 1777 return true; 1778 } 1779 1780 void DevirtModule::rebuildGlobal(VTableBits &B) { 1781 if (B.Before.Bytes.empty() && B.After.Bytes.empty()) 1782 return; 1783 1784 // Align the before byte array to the global's minimum alignment so that we 1785 // don't break any alignment requirements on the global. 1786 Align Alignment = M.getDataLayout().getValueOrABITypeAlignment( 1787 B.GV->getAlign(), B.GV->getValueType()); 1788 B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), Alignment)); 1789 1790 // Before was stored in reverse order; flip it now. 1791 for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I) 1792 std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]); 1793 1794 // Build an anonymous global containing the before bytes, followed by the 1795 // original initializer, followed by the after bytes. 1796 auto NewInit = ConstantStruct::getAnon( 1797 {ConstantDataArray::get(M.getContext(), B.Before.Bytes), 1798 B.GV->getInitializer(), 1799 ConstantDataArray::get(M.getContext(), B.After.Bytes)}); 1800 auto NewGV = 1801 new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(), 1802 GlobalVariable::PrivateLinkage, NewInit, "", B.GV); 1803 NewGV->setSection(B.GV->getSection()); 1804 NewGV->setComdat(B.GV->getComdat()); 1805 NewGV->setAlignment(B.GV->getAlign()); 1806 1807 // Copy the original vtable's metadata to the anonymous global, adjusting 1808 // offsets as required. 1809 NewGV->copyMetadata(B.GV, B.Before.Bytes.size()); 1810 1811 // Build an alias named after the original global, pointing at the second 1812 // element (the original initializer). 1813 auto Alias = GlobalAlias::create( 1814 B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "", 1815 ConstantExpr::getGetElementPtr( 1816 NewInit->getType(), NewGV, 1817 ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0), 1818 ConstantInt::get(Int32Ty, 1)}), 1819 &M); 1820 Alias->setVisibility(B.GV->getVisibility()); 1821 Alias->takeName(B.GV); 1822 1823 B.GV->replaceAllUsesWith(Alias); 1824 B.GV->eraseFromParent(); 1825 } 1826 1827 bool DevirtModule::areRemarksEnabled() { 1828 const auto &FL = M.getFunctionList(); 1829 for (const Function &Fn : FL) { 1830 const auto &BBL = Fn.getBasicBlockList(); 1831 if (BBL.empty()) 1832 continue; 1833 auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front()); 1834 return DI.isEnabled(); 1835 } 1836 return false; 1837 } 1838 1839 void DevirtModule::scanTypeTestUsers( 1840 Function *TypeTestFunc, 1841 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 1842 // Find all virtual calls via a virtual table pointer %p under an assumption 1843 // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p 1844 // points to a member of the type identifier %md. Group calls by (type ID, 1845 // offset) pair (effectively the identity of the virtual function) and store 1846 // to CallSlots. 1847 for (Use &U : llvm::make_early_inc_range(TypeTestFunc->uses())) { 1848 auto *CI = dyn_cast<CallInst>(U.getUser()); 1849 if (!CI) 1850 continue; 1851 1852 // Search for virtual calls based on %p and add them to DevirtCalls. 1853 SmallVector<DevirtCallSite, 1> DevirtCalls; 1854 SmallVector<CallInst *, 1> Assumes; 1855 auto &DT = LookupDomTree(*CI->getFunction()); 1856 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT); 1857 1858 Metadata *TypeId = 1859 cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata(); 1860 // If we found any, add them to CallSlots. 1861 if (!Assumes.empty()) { 1862 Value *Ptr = CI->getArgOperand(0)->stripPointerCasts(); 1863 for (DevirtCallSite Call : DevirtCalls) 1864 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, nullptr); 1865 } 1866 1867 auto RemoveTypeTestAssumes = [&]() { 1868 // We no longer need the assumes or the type test. 1869 for (auto Assume : Assumes) 1870 Assume->eraseFromParent(); 1871 // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we 1872 // may use the vtable argument later. 1873 if (CI->use_empty()) 1874 CI->eraseFromParent(); 1875 }; 1876 1877 // At this point we could remove all type test assume sequences, as they 1878 // were originally inserted for WPD. However, we can keep these in the 1879 // code stream for later analysis (e.g. to help drive more efficient ICP 1880 // sequences). They will eventually be removed by a second LowerTypeTests 1881 // invocation that cleans them up. In order to do this correctly, the first 1882 // LowerTypeTests invocation needs to know that they have "Unknown" type 1883 // test resolution, so that they aren't treated as Unsat and lowered to 1884 // False, which will break any uses on assumes. Below we remove any type 1885 // test assumes that will not be treated as Unknown by LTT. 1886 1887 // The type test assumes will be treated by LTT as Unsat if the type id is 1888 // not used on a global (in which case it has no entry in the TypeIdMap). 1889 if (!TypeIdMap.count(TypeId)) 1890 RemoveTypeTestAssumes(); 1891 1892 // For ThinLTO importing, we need to remove the type test assumes if this is 1893 // an MDString type id without a corresponding TypeIdSummary. Any 1894 // non-MDString type ids are ignored and treated as Unknown by LTT, so their 1895 // type test assumes can be kept. If the MDString type id is missing a 1896 // TypeIdSummary (e.g. because there was no use on a vcall, preventing the 1897 // exporting phase of WPD from analyzing it), then it would be treated as 1898 // Unsat by LTT and we need to remove its type test assumes here. If not 1899 // used on a vcall we don't need them for later optimization use in any 1900 // case. 1901 else if (ImportSummary && isa<MDString>(TypeId)) { 1902 const TypeIdSummary *TidSummary = 1903 ImportSummary->getTypeIdSummary(cast<MDString>(TypeId)->getString()); 1904 if (!TidSummary) 1905 RemoveTypeTestAssumes(); 1906 else 1907 // If one was created it should not be Unsat, because if we reached here 1908 // the type id was used on a global. 1909 assert(TidSummary->TTRes.TheKind != TypeTestResolution::Unsat); 1910 } 1911 } 1912 } 1913 1914 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) { 1915 Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test); 1916 1917 for (Use &U : llvm::make_early_inc_range(TypeCheckedLoadFunc->uses())) { 1918 auto *CI = dyn_cast<CallInst>(U.getUser()); 1919 if (!CI) 1920 continue; 1921 1922 Value *Ptr = CI->getArgOperand(0); 1923 Value *Offset = CI->getArgOperand(1); 1924 Value *TypeIdValue = CI->getArgOperand(2); 1925 Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata(); 1926 1927 SmallVector<DevirtCallSite, 1> DevirtCalls; 1928 SmallVector<Instruction *, 1> LoadedPtrs; 1929 SmallVector<Instruction *, 1> Preds; 1930 bool HasNonCallUses = false; 1931 auto &DT = LookupDomTree(*CI->getFunction()); 1932 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds, 1933 HasNonCallUses, CI, DT); 1934 1935 // Start by generating "pessimistic" code that explicitly loads the function 1936 // pointer from the vtable and performs the type check. If possible, we will 1937 // eliminate the load and the type check later. 1938 1939 // If possible, only generate the load at the point where it is used. 1940 // This helps avoid unnecessary spills. 1941 IRBuilder<> LoadB( 1942 (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI); 1943 Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset); 1944 Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy)); 1945 Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr); 1946 1947 for (Instruction *LoadedPtr : LoadedPtrs) { 1948 LoadedPtr->replaceAllUsesWith(LoadedValue); 1949 LoadedPtr->eraseFromParent(); 1950 } 1951 1952 // Likewise for the type test. 1953 IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI); 1954 CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue}); 1955 1956 for (Instruction *Pred : Preds) { 1957 Pred->replaceAllUsesWith(TypeTestCall); 1958 Pred->eraseFromParent(); 1959 } 1960 1961 // We have already erased any extractvalue instructions that refer to the 1962 // intrinsic call, but the intrinsic may have other non-extractvalue uses 1963 // (although this is unlikely). In that case, explicitly build a pair and 1964 // RAUW it. 1965 if (!CI->use_empty()) { 1966 Value *Pair = UndefValue::get(CI->getType()); 1967 IRBuilder<> B(CI); 1968 Pair = B.CreateInsertValue(Pair, LoadedValue, {0}); 1969 Pair = B.CreateInsertValue(Pair, TypeTestCall, {1}); 1970 CI->replaceAllUsesWith(Pair); 1971 } 1972 1973 // The number of unsafe uses is initially the number of uses. 1974 auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall]; 1975 NumUnsafeUses = DevirtCalls.size(); 1976 1977 // If the function pointer has a non-call user, we cannot eliminate the type 1978 // check, as one of those users may eventually call the pointer. Increment 1979 // the unsafe use count to make sure it cannot reach zero. 1980 if (HasNonCallUses) 1981 ++NumUnsafeUses; 1982 for (DevirtCallSite Call : DevirtCalls) { 1983 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, 1984 &NumUnsafeUses); 1985 } 1986 1987 CI->eraseFromParent(); 1988 } 1989 } 1990 1991 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) { 1992 auto *TypeId = dyn_cast<MDString>(Slot.TypeID); 1993 if (!TypeId) 1994 return; 1995 const TypeIdSummary *TidSummary = 1996 ImportSummary->getTypeIdSummary(TypeId->getString()); 1997 if (!TidSummary) 1998 return; 1999 auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset); 2000 if (ResI == TidSummary->WPDRes.end()) 2001 return; 2002 const WholeProgramDevirtResolution &Res = ResI->second; 2003 2004 if (Res.TheKind == WholeProgramDevirtResolution::SingleImpl) { 2005 assert(!Res.SingleImplName.empty()); 2006 // The type of the function in the declaration is irrelevant because every 2007 // call site will cast it to the correct type. 2008 Constant *SingleImpl = 2009 cast<Constant>(M.getOrInsertFunction(Res.SingleImplName, 2010 Type::getVoidTy(M.getContext())) 2011 .getCallee()); 2012 2013 // This is the import phase so we should not be exporting anything. 2014 bool IsExported = false; 2015 applySingleImplDevirt(SlotInfo, SingleImpl, IsExported); 2016 assert(!IsExported); 2017 } 2018 2019 for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) { 2020 auto I = Res.ResByArg.find(CSByConstantArg.first); 2021 if (I == Res.ResByArg.end()) 2022 continue; 2023 auto &ResByArg = I->second; 2024 // FIXME: We should figure out what to do about the "function name" argument 2025 // to the apply* functions, as the function names are unavailable during the 2026 // importing phase. For now we just pass the empty string. This does not 2027 // impact correctness because the function names are just used for remarks. 2028 switch (ResByArg.TheKind) { 2029 case WholeProgramDevirtResolution::ByArg::UniformRetVal: 2030 applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info); 2031 break; 2032 case WholeProgramDevirtResolution::ByArg::UniqueRetVal: { 2033 Constant *UniqueMemberAddr = 2034 importGlobal(Slot, CSByConstantArg.first, "unique_member"); 2035 applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info, 2036 UniqueMemberAddr); 2037 break; 2038 } 2039 case WholeProgramDevirtResolution::ByArg::VirtualConstProp: { 2040 Constant *Byte = importConstant(Slot, CSByConstantArg.first, "byte", 2041 Int32Ty, ResByArg.Byte); 2042 Constant *Bit = importConstant(Slot, CSByConstantArg.first, "bit", Int8Ty, 2043 ResByArg.Bit); 2044 applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit); 2045 break; 2046 } 2047 default: 2048 break; 2049 } 2050 } 2051 2052 if (Res.TheKind == WholeProgramDevirtResolution::BranchFunnel) { 2053 // The type of the function is irrelevant, because it's bitcast at calls 2054 // anyhow. 2055 Constant *JT = cast<Constant>( 2056 M.getOrInsertFunction(getGlobalName(Slot, {}, "branch_funnel"), 2057 Type::getVoidTy(M.getContext())) 2058 .getCallee()); 2059 bool IsExported = false; 2060 applyICallBranchFunnel(SlotInfo, JT, IsExported); 2061 assert(!IsExported); 2062 } 2063 } 2064 2065 void DevirtModule::removeRedundantTypeTests() { 2066 auto True = ConstantInt::getTrue(M.getContext()); 2067 for (auto &&U : NumUnsafeUsesForTypeTest) { 2068 if (U.second == 0) { 2069 U.first->replaceAllUsesWith(True); 2070 U.first->eraseFromParent(); 2071 } 2072 } 2073 } 2074 2075 ValueInfo 2076 DevirtModule::lookUpFunctionValueInfo(Function *TheFn, 2077 ModuleSummaryIndex *ExportSummary) { 2078 assert((ExportSummary != nullptr) && 2079 "Caller guarantees ExportSummary is not nullptr"); 2080 2081 const auto TheFnGUID = TheFn->getGUID(); 2082 const auto TheFnGUIDWithExportedName = GlobalValue::getGUID(TheFn->getName()); 2083 // Look up ValueInfo with the GUID in the current linkage. 2084 ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFnGUID); 2085 // If no entry is found and GUID is different from GUID computed using 2086 // exported name, look up ValueInfo with the exported name unconditionally. 2087 // This is a fallback. 2088 // 2089 // The reason to have a fallback: 2090 // 1. LTO could enable global value internalization via 2091 // `enable-lto-internalization`. 2092 // 2. The GUID in ExportedSummary is computed using exported name. 2093 if ((!TheFnVI) && (TheFnGUID != TheFnGUIDWithExportedName)) { 2094 TheFnVI = ExportSummary->getValueInfo(TheFnGUIDWithExportedName); 2095 } 2096 return TheFnVI; 2097 } 2098 2099 bool DevirtModule::mustBeUnreachableFunction( 2100 Function *const F, ModuleSummaryIndex *ExportSummary) { 2101 // First, learn unreachability by analyzing function IR. 2102 if (!F->isDeclaration()) { 2103 // A function must be unreachable if its entry block ends with an 2104 // 'unreachable'. 2105 return isa<UnreachableInst>(F->getEntryBlock().getTerminator()); 2106 } 2107 // Learn unreachability from ExportSummary if ExportSummary is present. 2108 return ExportSummary && 2109 ::mustBeUnreachableFunction( 2110 DevirtModule::lookUpFunctionValueInfo(F, ExportSummary)); 2111 } 2112 2113 bool DevirtModule::run() { 2114 // If only some of the modules were split, we cannot correctly perform 2115 // this transformation. We already checked for the presense of type tests 2116 // with partially split modules during the thin link, and would have emitted 2117 // an error if any were found, so here we can simply return. 2118 if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) || 2119 (ImportSummary && ImportSummary->partiallySplitLTOUnits())) 2120 return false; 2121 2122 Function *TypeTestFunc = 2123 M.getFunction(Intrinsic::getName(Intrinsic::type_test)); 2124 Function *TypeCheckedLoadFunc = 2125 M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load)); 2126 Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume)); 2127 2128 // Normally if there are no users of the devirtualization intrinsics in the 2129 // module, this pass has nothing to do. But if we are exporting, we also need 2130 // to handle any users that appear only in the function summaries. 2131 if (!ExportSummary && 2132 (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc || 2133 AssumeFunc->use_empty()) && 2134 (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty())) 2135 return false; 2136 2137 // Rebuild type metadata into a map for easy lookup. 2138 std::vector<VTableBits> Bits; 2139 DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap; 2140 buildTypeIdentifierMap(Bits, TypeIdMap); 2141 2142 if (TypeTestFunc && AssumeFunc) 2143 scanTypeTestUsers(TypeTestFunc, TypeIdMap); 2144 2145 if (TypeCheckedLoadFunc) 2146 scanTypeCheckedLoadUsers(TypeCheckedLoadFunc); 2147 2148 if (ImportSummary) { 2149 for (auto &S : CallSlots) 2150 importResolution(S.first, S.second); 2151 2152 removeRedundantTypeTests(); 2153 2154 // We have lowered or deleted the type instrinsics, so we will no 2155 // longer have enough information to reason about the liveness of virtual 2156 // function pointers in GlobalDCE. 2157 for (GlobalVariable &GV : M.globals()) 2158 GV.eraseMetadata(LLVMContext::MD_vcall_visibility); 2159 2160 // The rest of the code is only necessary when exporting or during regular 2161 // LTO, so we are done. 2162 return true; 2163 } 2164 2165 if (TypeIdMap.empty()) 2166 return true; 2167 2168 // Collect information from summary about which calls to try to devirtualize. 2169 if (ExportSummary) { 2170 DenseMap<GlobalValue::GUID, TinyPtrVector<Metadata *>> MetadataByGUID; 2171 for (auto &P : TypeIdMap) { 2172 if (auto *TypeId = dyn_cast<MDString>(P.first)) 2173 MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back( 2174 TypeId); 2175 } 2176 2177 for (auto &P : *ExportSummary) { 2178 for (auto &S : P.second.SummaryList) { 2179 auto *FS = dyn_cast<FunctionSummary>(S.get()); 2180 if (!FS) 2181 continue; 2182 // FIXME: Only add live functions. 2183 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 2184 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 2185 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 2186 } 2187 } 2188 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 2189 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 2190 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 2191 } 2192 } 2193 for (const FunctionSummary::ConstVCall &VC : 2194 FS->type_test_assume_const_vcalls()) { 2195 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 2196 CallSlots[{MD, VC.VFunc.Offset}] 2197 .ConstCSInfo[VC.Args] 2198 .addSummaryTypeTestAssumeUser(FS); 2199 } 2200 } 2201 for (const FunctionSummary::ConstVCall &VC : 2202 FS->type_checked_load_const_vcalls()) { 2203 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 2204 CallSlots[{MD, VC.VFunc.Offset}] 2205 .ConstCSInfo[VC.Args] 2206 .addSummaryTypeCheckedLoadUser(FS); 2207 } 2208 } 2209 } 2210 } 2211 } 2212 2213 // For each (type, offset) pair: 2214 bool DidVirtualConstProp = false; 2215 std::map<std::string, Function*> DevirtTargets; 2216 for (auto &S : CallSlots) { 2217 // Search each of the members of the type identifier for the virtual 2218 // function implementation at offset S.first.ByteOffset, and add to 2219 // TargetsForSlot. 2220 std::vector<VirtualCallTarget> TargetsForSlot; 2221 WholeProgramDevirtResolution *Res = nullptr; 2222 const std::set<TypeMemberInfo> &TypeMemberInfos = TypeIdMap[S.first.TypeID]; 2223 if (ExportSummary && isa<MDString>(S.first.TypeID) && 2224 TypeMemberInfos.size()) 2225 // For any type id used on a global's type metadata, create the type id 2226 // summary resolution regardless of whether we can devirtualize, so that 2227 // lower type tests knows the type id is not Unsat. If it was not used on 2228 // a global's type metadata, the TypeIdMap entry set will be empty, and 2229 // we don't want to create an entry (with the default Unknown type 2230 // resolution), which can prevent detection of the Unsat. 2231 Res = &ExportSummary 2232 ->getOrInsertTypeIdSummary( 2233 cast<MDString>(S.first.TypeID)->getString()) 2234 .WPDRes[S.first.ByteOffset]; 2235 if (tryFindVirtualCallTargets(TargetsForSlot, TypeMemberInfos, 2236 S.first.ByteOffset, ExportSummary)) { 2237 2238 if (!trySingleImplDevirt(ExportSummary, TargetsForSlot, S.second, Res)) { 2239 DidVirtualConstProp |= 2240 tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first); 2241 2242 tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first); 2243 } 2244 2245 // Collect functions devirtualized at least for one call site for stats. 2246 if (RemarksEnabled) 2247 for (const auto &T : TargetsForSlot) 2248 if (T.WasDevirt) 2249 DevirtTargets[std::string(T.Fn->getName())] = T.Fn; 2250 } 2251 2252 // CFI-specific: if we are exporting and any llvm.type.checked.load 2253 // intrinsics were *not* devirtualized, we need to add the resulting 2254 // llvm.type.test intrinsics to the function summaries so that the 2255 // LowerTypeTests pass will export them. 2256 if (ExportSummary && isa<MDString>(S.first.TypeID)) { 2257 auto GUID = 2258 GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString()); 2259 for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers) 2260 FS->addTypeTest(GUID); 2261 for (auto &CCS : S.second.ConstCSInfo) 2262 for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers) 2263 FS->addTypeTest(GUID); 2264 } 2265 } 2266 2267 if (RemarksEnabled) { 2268 // Generate remarks for each devirtualized function. 2269 for (const auto &DT : DevirtTargets) { 2270 Function *F = DT.second; 2271 2272 using namespace ore; 2273 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F) 2274 << "devirtualized " 2275 << NV("FunctionName", DT.first)); 2276 } 2277 } 2278 2279 removeRedundantTypeTests(); 2280 2281 // Rebuild each global we touched as part of virtual constant propagation to 2282 // include the before and after bytes. 2283 if (DidVirtualConstProp) 2284 for (VTableBits &B : Bits) 2285 rebuildGlobal(B); 2286 2287 // We have lowered or deleted the type instrinsics, so we will no 2288 // longer have enough information to reason about the liveness of virtual 2289 // function pointers in GlobalDCE. 2290 for (GlobalVariable &GV : M.globals()) 2291 GV.eraseMetadata(LLVMContext::MD_vcall_visibility); 2292 2293 return true; 2294 } 2295 2296 void DevirtIndex::run() { 2297 if (ExportSummary.typeIdCompatibleVtableMap().empty()) 2298 return; 2299 2300 DenseMap<GlobalValue::GUID, std::vector<StringRef>> NameByGUID; 2301 for (auto &P : ExportSummary.typeIdCompatibleVtableMap()) { 2302 NameByGUID[GlobalValue::getGUID(P.first)].push_back(P.first); 2303 } 2304 2305 // Collect information from summary about which calls to try to devirtualize. 2306 for (auto &P : ExportSummary) { 2307 for (auto &S : P.second.SummaryList) { 2308 auto *FS = dyn_cast<FunctionSummary>(S.get()); 2309 if (!FS) 2310 continue; 2311 // FIXME: Only add live functions. 2312 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 2313 for (StringRef Name : NameByGUID[VF.GUID]) { 2314 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 2315 } 2316 } 2317 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 2318 for (StringRef Name : NameByGUID[VF.GUID]) { 2319 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 2320 } 2321 } 2322 for (const FunctionSummary::ConstVCall &VC : 2323 FS->type_test_assume_const_vcalls()) { 2324 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 2325 CallSlots[{Name, VC.VFunc.Offset}] 2326 .ConstCSInfo[VC.Args] 2327 .addSummaryTypeTestAssumeUser(FS); 2328 } 2329 } 2330 for (const FunctionSummary::ConstVCall &VC : 2331 FS->type_checked_load_const_vcalls()) { 2332 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 2333 CallSlots[{Name, VC.VFunc.Offset}] 2334 .ConstCSInfo[VC.Args] 2335 .addSummaryTypeCheckedLoadUser(FS); 2336 } 2337 } 2338 } 2339 } 2340 2341 std::set<ValueInfo> DevirtTargets; 2342 // For each (type, offset) pair: 2343 for (auto &S : CallSlots) { 2344 // Search each of the members of the type identifier for the virtual 2345 // function implementation at offset S.first.ByteOffset, and add to 2346 // TargetsForSlot. 2347 std::vector<ValueInfo> TargetsForSlot; 2348 auto TidSummary = ExportSummary.getTypeIdCompatibleVtableSummary(S.first.TypeID); 2349 assert(TidSummary); 2350 // Create the type id summary resolution regardlness of whether we can 2351 // devirtualize, so that lower type tests knows the type id is used on 2352 // a global and not Unsat. 2353 WholeProgramDevirtResolution *Res = 2354 &ExportSummary.getOrInsertTypeIdSummary(S.first.TypeID) 2355 .WPDRes[S.first.ByteOffset]; 2356 if (tryFindVirtualCallTargets(TargetsForSlot, *TidSummary, 2357 S.first.ByteOffset)) { 2358 2359 if (!trySingleImplDevirt(TargetsForSlot, S.first, S.second, Res, 2360 DevirtTargets)) 2361 continue; 2362 } 2363 } 2364 2365 // Optionally have the thin link print message for each devirtualized 2366 // function. 2367 if (PrintSummaryDevirt) 2368 for (const auto &DT : DevirtTargets) 2369 errs() << "Devirtualized call to " << DT << "\n"; 2370 } 2371