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 // Don't upgrade the visibility for symbols exported to the dynamic 870 // linker, as we have no information on their eventual use. 871 if (DynamicExportSymbols.count(P.first)) 872 continue; 873 for (auto &S : P.second.SummaryList) { 874 auto *GVar = dyn_cast<GlobalVarSummary>(S.get()); 875 if (!GVar || 876 GVar->getVCallVisibility() != GlobalObject::VCallVisibilityPublic) 877 continue; 878 GVar->setVCallVisibility(GlobalObject::VCallVisibilityLinkageUnit); 879 } 880 } 881 } 882 883 void runWholeProgramDevirtOnIndex( 884 ModuleSummaryIndex &Summary, std::set<GlobalValue::GUID> &ExportedGUIDs, 885 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 886 DevirtIndex(Summary, ExportedGUIDs, LocalWPDTargetsMap).run(); 887 } 888 889 void updateIndexWPDForExports( 890 ModuleSummaryIndex &Summary, 891 function_ref<bool(StringRef, ValueInfo)> isExported, 892 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 893 for (auto &T : LocalWPDTargetsMap) { 894 auto &VI = T.first; 895 // This was enforced earlier during trySingleImplDevirt. 896 assert(VI.getSummaryList().size() == 1 && 897 "Devirt of local target has more than one copy"); 898 auto &S = VI.getSummaryList()[0]; 899 if (!isExported(S->modulePath(), VI)) 900 continue; 901 902 // It's been exported by a cross module import. 903 for (auto &SlotSummary : T.second) { 904 auto *TIdSum = Summary.getTypeIdSummary(SlotSummary.TypeID); 905 assert(TIdSum); 906 auto WPDRes = TIdSum->WPDRes.find(SlotSummary.ByteOffset); 907 assert(WPDRes != TIdSum->WPDRes.end()); 908 WPDRes->second.SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 909 WPDRes->second.SingleImplName, 910 Summary.getModuleHash(S->modulePath())); 911 } 912 } 913 } 914 915 } // end namespace llvm 916 917 static Error checkCombinedSummaryForTesting(ModuleSummaryIndex *Summary) { 918 // Check that summary index contains regular LTO module when performing 919 // export to prevent occasional use of index from pure ThinLTO compilation 920 // (-fno-split-lto-module). This kind of summary index is passed to 921 // DevirtIndex::run, not to DevirtModule::run used by opt/runForTesting. 922 const auto &ModPaths = Summary->modulePaths(); 923 if (ClSummaryAction != PassSummaryAction::Import && 924 ModPaths.find(ModuleSummaryIndex::getRegularLTOModuleName()) == 925 ModPaths.end()) 926 return createStringError( 927 errc::invalid_argument, 928 "combined summary should contain Regular LTO module"); 929 return ErrorSuccess(); 930 } 931 932 bool DevirtModule::runForTesting( 933 Module &M, function_ref<AAResults &(Function &)> AARGetter, 934 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 935 function_ref<DominatorTree &(Function &)> LookupDomTree) { 936 std::unique_ptr<ModuleSummaryIndex> Summary = 937 std::make_unique<ModuleSummaryIndex>(/*HaveGVs=*/false); 938 939 // Handle the command-line summary arguments. This code is for testing 940 // purposes only, so we handle errors directly. 941 if (!ClReadSummary.empty()) { 942 ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary + 943 ": "); 944 auto ReadSummaryFile = 945 ExitOnErr(errorOrToExpected(MemoryBuffer::getFile(ClReadSummary))); 946 if (Expected<std::unique_ptr<ModuleSummaryIndex>> SummaryOrErr = 947 getModuleSummaryIndex(*ReadSummaryFile)) { 948 Summary = std::move(*SummaryOrErr); 949 ExitOnErr(checkCombinedSummaryForTesting(Summary.get())); 950 } else { 951 // Try YAML if we've failed with bitcode. 952 consumeError(SummaryOrErr.takeError()); 953 yaml::Input In(ReadSummaryFile->getBuffer()); 954 In >> *Summary; 955 ExitOnErr(errorCodeToError(In.error())); 956 } 957 } 958 959 bool Changed = 960 DevirtModule(M, AARGetter, OREGetter, LookupDomTree, 961 ClSummaryAction == PassSummaryAction::Export ? Summary.get() 962 : nullptr, 963 ClSummaryAction == PassSummaryAction::Import ? Summary.get() 964 : nullptr) 965 .run(); 966 967 if (!ClWriteSummary.empty()) { 968 ExitOnError ExitOnErr( 969 "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": "); 970 std::error_code EC; 971 if (StringRef(ClWriteSummary).endswith(".bc")) { 972 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_None); 973 ExitOnErr(errorCodeToError(EC)); 974 writeIndexToFile(*Summary, OS); 975 } else { 976 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_TextWithCRLF); 977 ExitOnErr(errorCodeToError(EC)); 978 yaml::Output Out(OS); 979 Out << *Summary; 980 } 981 } 982 983 return Changed; 984 } 985 986 void DevirtModule::buildTypeIdentifierMap( 987 std::vector<VTableBits> &Bits, 988 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 989 DenseMap<GlobalVariable *, VTableBits *> GVToBits; 990 Bits.reserve(M.getGlobalList().size()); 991 SmallVector<MDNode *, 2> Types; 992 for (GlobalVariable &GV : M.globals()) { 993 Types.clear(); 994 GV.getMetadata(LLVMContext::MD_type, Types); 995 if (GV.isDeclaration() || Types.empty()) 996 continue; 997 998 VTableBits *&BitsPtr = GVToBits[&GV]; 999 if (!BitsPtr) { 1000 Bits.emplace_back(); 1001 Bits.back().GV = &GV; 1002 Bits.back().ObjectSize = 1003 M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType()); 1004 BitsPtr = &Bits.back(); 1005 } 1006 1007 for (MDNode *Type : Types) { 1008 auto TypeID = Type->getOperand(1).get(); 1009 1010 uint64_t Offset = 1011 cast<ConstantInt>( 1012 cast<ConstantAsMetadata>(Type->getOperand(0))->getValue()) 1013 ->getZExtValue(); 1014 1015 TypeIdMap[TypeID].insert({BitsPtr, Offset}); 1016 } 1017 } 1018 } 1019 1020 bool DevirtModule::tryFindVirtualCallTargets( 1021 std::vector<VirtualCallTarget> &TargetsForSlot, 1022 const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset, 1023 ModuleSummaryIndex *ExportSummary) { 1024 for (const TypeMemberInfo &TM : TypeMemberInfos) { 1025 if (!TM.Bits->GV->isConstant()) 1026 return false; 1027 1028 // We cannot perform whole program devirtualization analysis on a vtable 1029 // with public LTO visibility. 1030 if (TM.Bits->GV->getVCallVisibility() == 1031 GlobalObject::VCallVisibilityPublic) 1032 return false; 1033 1034 Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(), 1035 TM.Offset + ByteOffset, M); 1036 if (!Ptr) 1037 return false; 1038 1039 auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts()); 1040 if (!Fn) 1041 return false; 1042 1043 if (FunctionsToSkip.match(Fn->getName())) 1044 return false; 1045 1046 // We can disregard __cxa_pure_virtual as a possible call target, as 1047 // calls to pure virtuals are UB. 1048 if (Fn->getName() == "__cxa_pure_virtual") 1049 continue; 1050 1051 // We can disregard unreachable functions as possible call targets, as 1052 // unreachable functions shouldn't be called. 1053 if (mustBeUnreachableFunction(Fn, ExportSummary)) 1054 continue; 1055 1056 TargetsForSlot.push_back({Fn, &TM}); 1057 } 1058 1059 // Give up if we couldn't find any targets. 1060 return !TargetsForSlot.empty(); 1061 } 1062 1063 bool DevirtIndex::tryFindVirtualCallTargets( 1064 std::vector<ValueInfo> &TargetsForSlot, const TypeIdCompatibleVtableInfo TIdInfo, 1065 uint64_t ByteOffset) { 1066 for (const TypeIdOffsetVtableInfo &P : TIdInfo) { 1067 // Find a representative copy of the vtable initializer. 1068 // We can have multiple available_externally, linkonce_odr and weak_odr 1069 // vtable initializers. We can also have multiple external vtable 1070 // initializers in the case of comdats, which we cannot check here. 1071 // The linker should give an error in this case. 1072 // 1073 // Also, handle the case of same-named local Vtables with the same path 1074 // and therefore the same GUID. This can happen if there isn't enough 1075 // distinguishing path when compiling the source file. In that case we 1076 // conservatively return false early. 1077 const GlobalVarSummary *VS = nullptr; 1078 bool LocalFound = false; 1079 for (auto &S : P.VTableVI.getSummaryList()) { 1080 if (GlobalValue::isLocalLinkage(S->linkage())) { 1081 if (LocalFound) 1082 return false; 1083 LocalFound = true; 1084 } 1085 auto *CurVS = cast<GlobalVarSummary>(S->getBaseObject()); 1086 if (!CurVS->vTableFuncs().empty() || 1087 // Previously clang did not attach the necessary type metadata to 1088 // available_externally vtables, in which case there would not 1089 // be any vtable functions listed in the summary and we need 1090 // to treat this case conservatively (in case the bitcode is old). 1091 // However, we will also not have any vtable functions in the 1092 // case of a pure virtual base class. In that case we do want 1093 // to set VS to avoid treating it conservatively. 1094 !GlobalValue::isAvailableExternallyLinkage(S->linkage())) { 1095 VS = CurVS; 1096 // We cannot perform whole program devirtualization analysis on a vtable 1097 // with public LTO visibility. 1098 if (VS->getVCallVisibility() == GlobalObject::VCallVisibilityPublic) 1099 return false; 1100 } 1101 } 1102 // There will be no VS if all copies are available_externally having no 1103 // type metadata. In that case we can't safely perform WPD. 1104 if (!VS) 1105 return false; 1106 if (!VS->isLive()) 1107 continue; 1108 for (auto VTP : VS->vTableFuncs()) { 1109 if (VTP.VTableOffset != P.AddressPointOffset + ByteOffset) 1110 continue; 1111 1112 if (mustBeUnreachableFunction(VTP.FuncVI)) 1113 continue; 1114 1115 TargetsForSlot.push_back(VTP.FuncVI); 1116 } 1117 } 1118 1119 // Give up if we couldn't find any targets. 1120 return !TargetsForSlot.empty(); 1121 } 1122 1123 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo, 1124 Constant *TheFn, bool &IsExported) { 1125 // Don't devirtualize function if we're told to skip it 1126 // in -wholeprogramdevirt-skip. 1127 if (FunctionsToSkip.match(TheFn->stripPointerCasts()->getName())) 1128 return; 1129 auto Apply = [&](CallSiteInfo &CSInfo) { 1130 for (auto &&VCallSite : CSInfo.CallSites) { 1131 if (!OptimizedCalls.insert(&VCallSite.CB).second) 1132 continue; 1133 1134 if (RemarksEnabled) 1135 VCallSite.emitRemark("single-impl", 1136 TheFn->stripPointerCasts()->getName(), OREGetter); 1137 auto &CB = VCallSite.CB; 1138 assert(!CB.getCalledFunction() && "devirtualizing direct call?"); 1139 IRBuilder<> Builder(&CB); 1140 Value *Callee = 1141 Builder.CreateBitCast(TheFn, CB.getCalledOperand()->getType()); 1142 1143 // If checking is enabled, add support to compare the virtual function 1144 // pointer to the devirtualized target. In case of a mismatch, perform a 1145 // debug trap. 1146 if (CheckDevirt) { 1147 auto *Cond = Builder.CreateICmpNE(CB.getCalledOperand(), Callee); 1148 Instruction *ThenTerm = 1149 SplitBlockAndInsertIfThen(Cond, &CB, /*Unreachable=*/false); 1150 Builder.SetInsertPoint(ThenTerm); 1151 Function *TrapFn = Intrinsic::getDeclaration(&M, Intrinsic::debugtrap); 1152 auto *CallTrap = Builder.CreateCall(TrapFn); 1153 CallTrap->setDebugLoc(CB.getDebugLoc()); 1154 } 1155 1156 // Devirtualize. 1157 CB.setCalledOperand(Callee); 1158 1159 // This use is no longer unsafe. 1160 if (VCallSite.NumUnsafeUses) 1161 --*VCallSite.NumUnsafeUses; 1162 } 1163 if (CSInfo.isExported()) 1164 IsExported = true; 1165 CSInfo.markDevirt(); 1166 }; 1167 Apply(SlotInfo.CSInfo); 1168 for (auto &P : SlotInfo.ConstCSInfo) 1169 Apply(P.second); 1170 } 1171 1172 static bool AddCalls(VTableSlotInfo &SlotInfo, const ValueInfo &Callee) { 1173 // We can't add calls if we haven't seen a definition 1174 if (Callee.getSummaryList().empty()) 1175 return false; 1176 1177 // Insert calls into the summary index so that the devirtualized targets 1178 // are eligible for import. 1179 // FIXME: Annotate type tests with hotness. For now, mark these as hot 1180 // to better ensure we have the opportunity to inline them. 1181 bool IsExported = false; 1182 auto &S = Callee.getSummaryList()[0]; 1183 CalleeInfo CI(CalleeInfo::HotnessType::Hot, /* RelBF = */ 0); 1184 auto AddCalls = [&](CallSiteInfo &CSInfo) { 1185 for (auto *FS : CSInfo.SummaryTypeCheckedLoadUsers) { 1186 FS->addCall({Callee, CI}); 1187 IsExported |= S->modulePath() != FS->modulePath(); 1188 } 1189 for (auto *FS : CSInfo.SummaryTypeTestAssumeUsers) { 1190 FS->addCall({Callee, CI}); 1191 IsExported |= S->modulePath() != FS->modulePath(); 1192 } 1193 }; 1194 AddCalls(SlotInfo.CSInfo); 1195 for (auto &P : SlotInfo.ConstCSInfo) 1196 AddCalls(P.second); 1197 return IsExported; 1198 } 1199 1200 bool DevirtModule::trySingleImplDevirt( 1201 ModuleSummaryIndex *ExportSummary, 1202 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1203 WholeProgramDevirtResolution *Res) { 1204 // See if the program contains a single implementation of this virtual 1205 // function. 1206 Function *TheFn = TargetsForSlot[0].Fn; 1207 for (auto &&Target : TargetsForSlot) 1208 if (TheFn != Target.Fn) 1209 return false; 1210 1211 // If so, update each call site to call that implementation directly. 1212 if (RemarksEnabled) 1213 TargetsForSlot[0].WasDevirt = true; 1214 1215 bool IsExported = false; 1216 applySingleImplDevirt(SlotInfo, TheFn, IsExported); 1217 if (!IsExported) 1218 return false; 1219 1220 // If the only implementation has local linkage, we must promote to external 1221 // to make it visible to thin LTO objects. We can only get here during the 1222 // ThinLTO export phase. 1223 if (TheFn->hasLocalLinkage()) { 1224 std::string NewName = (TheFn->getName() + ".llvm.merged").str(); 1225 1226 // Since we are renaming the function, any comdats with the same name must 1227 // also be renamed. This is required when targeting COFF, as the comdat name 1228 // must match one of the names of the symbols in the comdat. 1229 if (Comdat *C = TheFn->getComdat()) { 1230 if (C->getName() == TheFn->getName()) { 1231 Comdat *NewC = M.getOrInsertComdat(NewName); 1232 NewC->setSelectionKind(C->getSelectionKind()); 1233 for (GlobalObject &GO : M.global_objects()) 1234 if (GO.getComdat() == C) 1235 GO.setComdat(NewC); 1236 } 1237 } 1238 1239 TheFn->setLinkage(GlobalValue::ExternalLinkage); 1240 TheFn->setVisibility(GlobalValue::HiddenVisibility); 1241 TheFn->setName(NewName); 1242 } 1243 if (ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFn->getGUID())) 1244 // Any needed promotion of 'TheFn' has already been done during 1245 // LTO unit split, so we can ignore return value of AddCalls. 1246 AddCalls(SlotInfo, TheFnVI); 1247 1248 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1249 Res->SingleImplName = std::string(TheFn->getName()); 1250 1251 return true; 1252 } 1253 1254 bool DevirtIndex::trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 1255 VTableSlotSummary &SlotSummary, 1256 VTableSlotInfo &SlotInfo, 1257 WholeProgramDevirtResolution *Res, 1258 std::set<ValueInfo> &DevirtTargets) { 1259 // See if the program contains a single implementation of this virtual 1260 // function. 1261 auto TheFn = TargetsForSlot[0]; 1262 for (auto &&Target : TargetsForSlot) 1263 if (TheFn != Target) 1264 return false; 1265 1266 // Don't devirtualize if we don't have target definition. 1267 auto Size = TheFn.getSummaryList().size(); 1268 if (!Size) 1269 return false; 1270 1271 // Don't devirtualize function if we're told to skip it 1272 // in -wholeprogramdevirt-skip. 1273 if (FunctionsToSkip.match(TheFn.name())) 1274 return false; 1275 1276 // If the summary list contains multiple summaries where at least one is 1277 // a local, give up, as we won't know which (possibly promoted) name to use. 1278 for (auto &S : TheFn.getSummaryList()) 1279 if (GlobalValue::isLocalLinkage(S->linkage()) && Size > 1) 1280 return false; 1281 1282 // Collect functions devirtualized at least for one call site for stats. 1283 if (PrintSummaryDevirt) 1284 DevirtTargets.insert(TheFn); 1285 1286 auto &S = TheFn.getSummaryList()[0]; 1287 bool IsExported = AddCalls(SlotInfo, TheFn); 1288 if (IsExported) 1289 ExportedGUIDs.insert(TheFn.getGUID()); 1290 1291 // Record in summary for use in devirtualization during the ThinLTO import 1292 // step. 1293 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1294 if (GlobalValue::isLocalLinkage(S->linkage())) { 1295 if (IsExported) 1296 // If target is a local function and we are exporting it by 1297 // devirtualizing a call in another module, we need to record the 1298 // promoted name. 1299 Res->SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 1300 TheFn.name(), ExportSummary.getModuleHash(S->modulePath())); 1301 else { 1302 LocalWPDTargetsMap[TheFn].push_back(SlotSummary); 1303 Res->SingleImplName = std::string(TheFn.name()); 1304 } 1305 } else 1306 Res->SingleImplName = std::string(TheFn.name()); 1307 1308 // Name will be empty if this thin link driven off of serialized combined 1309 // index (e.g. llvm-lto). However, WPD is not supported/invoked for the 1310 // legacy LTO API anyway. 1311 assert(!Res->SingleImplName.empty()); 1312 1313 return true; 1314 } 1315 1316 void DevirtModule::tryICallBranchFunnel( 1317 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1318 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1319 Triple T(M.getTargetTriple()); 1320 if (T.getArch() != Triple::x86_64) 1321 return; 1322 1323 if (TargetsForSlot.size() > ClThreshold) 1324 return; 1325 1326 bool HasNonDevirt = !SlotInfo.CSInfo.AllCallSitesDevirted; 1327 if (!HasNonDevirt) 1328 for (auto &P : SlotInfo.ConstCSInfo) 1329 if (!P.second.AllCallSitesDevirted) { 1330 HasNonDevirt = true; 1331 break; 1332 } 1333 1334 if (!HasNonDevirt) 1335 return; 1336 1337 FunctionType *FT = 1338 FunctionType::get(Type::getVoidTy(M.getContext()), {Int8PtrTy}, true); 1339 Function *JT; 1340 if (isa<MDString>(Slot.TypeID)) { 1341 JT = Function::Create(FT, Function::ExternalLinkage, 1342 M.getDataLayout().getProgramAddressSpace(), 1343 getGlobalName(Slot, {}, "branch_funnel"), &M); 1344 JT->setVisibility(GlobalValue::HiddenVisibility); 1345 } else { 1346 JT = Function::Create(FT, Function::InternalLinkage, 1347 M.getDataLayout().getProgramAddressSpace(), 1348 "branch_funnel", &M); 1349 } 1350 JT->addParamAttr(0, Attribute::Nest); 1351 1352 std::vector<Value *> JTArgs; 1353 JTArgs.push_back(JT->arg_begin()); 1354 for (auto &T : TargetsForSlot) { 1355 JTArgs.push_back(getMemberAddr(T.TM)); 1356 JTArgs.push_back(T.Fn); 1357 } 1358 1359 BasicBlock *BB = BasicBlock::Create(M.getContext(), "", JT, nullptr); 1360 Function *Intr = 1361 Intrinsic::getDeclaration(&M, llvm::Intrinsic::icall_branch_funnel, {}); 1362 1363 auto *CI = CallInst::Create(Intr, JTArgs, "", BB); 1364 CI->setTailCallKind(CallInst::TCK_MustTail); 1365 ReturnInst::Create(M.getContext(), nullptr, BB); 1366 1367 bool IsExported = false; 1368 applyICallBranchFunnel(SlotInfo, JT, IsExported); 1369 if (IsExported) 1370 Res->TheKind = WholeProgramDevirtResolution::BranchFunnel; 1371 } 1372 1373 void DevirtModule::applyICallBranchFunnel(VTableSlotInfo &SlotInfo, 1374 Constant *JT, bool &IsExported) { 1375 auto Apply = [&](CallSiteInfo &CSInfo) { 1376 if (CSInfo.isExported()) 1377 IsExported = true; 1378 if (CSInfo.AllCallSitesDevirted) 1379 return; 1380 for (auto &&VCallSite : CSInfo.CallSites) { 1381 CallBase &CB = VCallSite.CB; 1382 1383 // Jump tables are only profitable if the retpoline mitigation is enabled. 1384 Attribute FSAttr = CB.getCaller()->getFnAttribute("target-features"); 1385 if (!FSAttr.isValid() || 1386 !FSAttr.getValueAsString().contains("+retpoline")) 1387 continue; 1388 1389 if (RemarksEnabled) 1390 VCallSite.emitRemark("branch-funnel", 1391 JT->stripPointerCasts()->getName(), OREGetter); 1392 1393 // Pass the address of the vtable in the nest register, which is r10 on 1394 // x86_64. 1395 std::vector<Type *> NewArgs; 1396 NewArgs.push_back(Int8PtrTy); 1397 append_range(NewArgs, CB.getFunctionType()->params()); 1398 FunctionType *NewFT = 1399 FunctionType::get(CB.getFunctionType()->getReturnType(), NewArgs, 1400 CB.getFunctionType()->isVarArg()); 1401 PointerType *NewFTPtr = PointerType::getUnqual(NewFT); 1402 1403 IRBuilder<> IRB(&CB); 1404 std::vector<Value *> Args; 1405 Args.push_back(IRB.CreateBitCast(VCallSite.VTable, Int8PtrTy)); 1406 llvm::append_range(Args, CB.args()); 1407 1408 CallBase *NewCS = nullptr; 1409 if (isa<CallInst>(CB)) 1410 NewCS = IRB.CreateCall(NewFT, IRB.CreateBitCast(JT, NewFTPtr), Args); 1411 else 1412 NewCS = IRB.CreateInvoke(NewFT, IRB.CreateBitCast(JT, NewFTPtr), 1413 cast<InvokeInst>(CB).getNormalDest(), 1414 cast<InvokeInst>(CB).getUnwindDest(), Args); 1415 NewCS->setCallingConv(CB.getCallingConv()); 1416 1417 AttributeList Attrs = CB.getAttributes(); 1418 std::vector<AttributeSet> NewArgAttrs; 1419 NewArgAttrs.push_back(AttributeSet::get( 1420 M.getContext(), ArrayRef<Attribute>{Attribute::get( 1421 M.getContext(), Attribute::Nest)})); 1422 for (unsigned I = 0; I + 2 < Attrs.getNumAttrSets(); ++I) 1423 NewArgAttrs.push_back(Attrs.getParamAttrs(I)); 1424 NewCS->setAttributes( 1425 AttributeList::get(M.getContext(), Attrs.getFnAttrs(), 1426 Attrs.getRetAttrs(), NewArgAttrs)); 1427 1428 CB.replaceAllUsesWith(NewCS); 1429 CB.eraseFromParent(); 1430 1431 // This use is no longer unsafe. 1432 if (VCallSite.NumUnsafeUses) 1433 --*VCallSite.NumUnsafeUses; 1434 } 1435 // Don't mark as devirtualized because there may be callers compiled without 1436 // retpoline mitigation, which would mean that they are lowered to 1437 // llvm.type.test and therefore require an llvm.type.test resolution for the 1438 // type identifier. 1439 }; 1440 Apply(SlotInfo.CSInfo); 1441 for (auto &P : SlotInfo.ConstCSInfo) 1442 Apply(P.second); 1443 } 1444 1445 bool DevirtModule::tryEvaluateFunctionsWithArgs( 1446 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1447 ArrayRef<uint64_t> Args) { 1448 // Evaluate each function and store the result in each target's RetVal 1449 // field. 1450 for (VirtualCallTarget &Target : TargetsForSlot) { 1451 if (Target.Fn->arg_size() != Args.size() + 1) 1452 return false; 1453 1454 Evaluator Eval(M.getDataLayout(), nullptr); 1455 SmallVector<Constant *, 2> EvalArgs; 1456 EvalArgs.push_back( 1457 Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0))); 1458 for (unsigned I = 0; I != Args.size(); ++I) { 1459 auto *ArgTy = dyn_cast<IntegerType>( 1460 Target.Fn->getFunctionType()->getParamType(I + 1)); 1461 if (!ArgTy) 1462 return false; 1463 EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I])); 1464 } 1465 1466 Constant *RetVal; 1467 if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) || 1468 !isa<ConstantInt>(RetVal)) 1469 return false; 1470 Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue(); 1471 } 1472 return true; 1473 } 1474 1475 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1476 uint64_t TheRetVal) { 1477 for (auto Call : CSInfo.CallSites) { 1478 if (!OptimizedCalls.insert(&Call.CB).second) 1479 continue; 1480 Call.replaceAndErase( 1481 "uniform-ret-val", FnName, RemarksEnabled, OREGetter, 1482 ConstantInt::get(cast<IntegerType>(Call.CB.getType()), TheRetVal)); 1483 } 1484 CSInfo.markDevirt(); 1485 } 1486 1487 bool DevirtModule::tryUniformRetValOpt( 1488 MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo, 1489 WholeProgramDevirtResolution::ByArg *Res) { 1490 // Uniform return value optimization. If all functions return the same 1491 // constant, replace all calls with that constant. 1492 uint64_t TheRetVal = TargetsForSlot[0].RetVal; 1493 for (const VirtualCallTarget &Target : TargetsForSlot) 1494 if (Target.RetVal != TheRetVal) 1495 return false; 1496 1497 if (CSInfo.isExported()) { 1498 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniformRetVal; 1499 Res->Info = TheRetVal; 1500 } 1501 1502 applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal); 1503 if (RemarksEnabled) 1504 for (auto &&Target : TargetsForSlot) 1505 Target.WasDevirt = true; 1506 return true; 1507 } 1508 1509 std::string DevirtModule::getGlobalName(VTableSlot Slot, 1510 ArrayRef<uint64_t> Args, 1511 StringRef Name) { 1512 std::string FullName = "__typeid_"; 1513 raw_string_ostream OS(FullName); 1514 OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset; 1515 for (uint64_t Arg : Args) 1516 OS << '_' << Arg; 1517 OS << '_' << Name; 1518 return OS.str(); 1519 } 1520 1521 bool DevirtModule::shouldExportConstantsAsAbsoluteSymbols() { 1522 Triple T(M.getTargetTriple()); 1523 return T.isX86() && T.getObjectFormat() == Triple::ELF; 1524 } 1525 1526 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1527 StringRef Name, Constant *C) { 1528 GlobalAlias *GA = GlobalAlias::create(Int8Ty, 0, GlobalValue::ExternalLinkage, 1529 getGlobalName(Slot, Args, Name), C, &M); 1530 GA->setVisibility(GlobalValue::HiddenVisibility); 1531 } 1532 1533 void DevirtModule::exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1534 StringRef Name, uint32_t Const, 1535 uint32_t &Storage) { 1536 if (shouldExportConstantsAsAbsoluteSymbols()) { 1537 exportGlobal( 1538 Slot, Args, Name, 1539 ConstantExpr::getIntToPtr(ConstantInt::get(Int32Ty, Const), Int8PtrTy)); 1540 return; 1541 } 1542 1543 Storage = Const; 1544 } 1545 1546 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1547 StringRef Name) { 1548 Constant *C = 1549 M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Arr0Ty); 1550 auto *GV = dyn_cast<GlobalVariable>(C); 1551 if (GV) 1552 GV->setVisibility(GlobalValue::HiddenVisibility); 1553 return C; 1554 } 1555 1556 Constant *DevirtModule::importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1557 StringRef Name, IntegerType *IntTy, 1558 uint32_t Storage) { 1559 if (!shouldExportConstantsAsAbsoluteSymbols()) 1560 return ConstantInt::get(IntTy, Storage); 1561 1562 Constant *C = importGlobal(Slot, Args, Name); 1563 auto *GV = cast<GlobalVariable>(C->stripPointerCasts()); 1564 C = ConstantExpr::getPtrToInt(C, IntTy); 1565 1566 // We only need to set metadata if the global is newly created, in which 1567 // case it would not have hidden visibility. 1568 if (GV->hasMetadata(LLVMContext::MD_absolute_symbol)) 1569 return C; 1570 1571 auto SetAbsRange = [&](uint64_t Min, uint64_t Max) { 1572 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min)); 1573 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max)); 1574 GV->setMetadata(LLVMContext::MD_absolute_symbol, 1575 MDNode::get(M.getContext(), {MinC, MaxC})); 1576 }; 1577 unsigned AbsWidth = IntTy->getBitWidth(); 1578 if (AbsWidth == IntPtrTy->getBitWidth()) 1579 SetAbsRange(~0ull, ~0ull); // Full set. 1580 else 1581 SetAbsRange(0, 1ull << AbsWidth); 1582 return C; 1583 } 1584 1585 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1586 bool IsOne, 1587 Constant *UniqueMemberAddr) { 1588 for (auto &&Call : CSInfo.CallSites) { 1589 if (!OptimizedCalls.insert(&Call.CB).second) 1590 continue; 1591 IRBuilder<> B(&Call.CB); 1592 Value *Cmp = 1593 B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE, Call.VTable, 1594 B.CreateBitCast(UniqueMemberAddr, Call.VTable->getType())); 1595 Cmp = B.CreateZExt(Cmp, Call.CB.getType()); 1596 Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter, 1597 Cmp); 1598 } 1599 CSInfo.markDevirt(); 1600 } 1601 1602 Constant *DevirtModule::getMemberAddr(const TypeMemberInfo *M) { 1603 Constant *C = ConstantExpr::getBitCast(M->Bits->GV, Int8PtrTy); 1604 return ConstantExpr::getGetElementPtr(Int8Ty, C, 1605 ConstantInt::get(Int64Ty, M->Offset)); 1606 } 1607 1608 bool DevirtModule::tryUniqueRetValOpt( 1609 unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1610 CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res, 1611 VTableSlot Slot, ArrayRef<uint64_t> Args) { 1612 // IsOne controls whether we look for a 0 or a 1. 1613 auto tryUniqueRetValOptFor = [&](bool IsOne) { 1614 const TypeMemberInfo *UniqueMember = nullptr; 1615 for (const VirtualCallTarget &Target : TargetsForSlot) { 1616 if (Target.RetVal == (IsOne ? 1 : 0)) { 1617 if (UniqueMember) 1618 return false; 1619 UniqueMember = Target.TM; 1620 } 1621 } 1622 1623 // We should have found a unique member or bailed out by now. We already 1624 // checked for a uniform return value in tryUniformRetValOpt. 1625 assert(UniqueMember); 1626 1627 Constant *UniqueMemberAddr = getMemberAddr(UniqueMember); 1628 if (CSInfo.isExported()) { 1629 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniqueRetVal; 1630 Res->Info = IsOne; 1631 1632 exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr); 1633 } 1634 1635 // Replace each call with the comparison. 1636 applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne, 1637 UniqueMemberAddr); 1638 1639 // Update devirtualization statistics for targets. 1640 if (RemarksEnabled) 1641 for (auto &&Target : TargetsForSlot) 1642 Target.WasDevirt = true; 1643 1644 return true; 1645 }; 1646 1647 if (BitWidth == 1) { 1648 if (tryUniqueRetValOptFor(true)) 1649 return true; 1650 if (tryUniqueRetValOptFor(false)) 1651 return true; 1652 } 1653 return false; 1654 } 1655 1656 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 1657 Constant *Byte, Constant *Bit) { 1658 for (auto Call : CSInfo.CallSites) { 1659 if (!OptimizedCalls.insert(&Call.CB).second) 1660 continue; 1661 auto *RetType = cast<IntegerType>(Call.CB.getType()); 1662 IRBuilder<> B(&Call.CB); 1663 Value *Addr = 1664 B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte); 1665 if (RetType->getBitWidth() == 1) { 1666 Value *Bits = B.CreateLoad(Int8Ty, Addr); 1667 Value *BitsAndBit = B.CreateAnd(Bits, Bit); 1668 auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0)); 1669 Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled, 1670 OREGetter, IsBitSet); 1671 } else { 1672 Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo()); 1673 Value *Val = B.CreateLoad(RetType, ValAddr); 1674 Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled, 1675 OREGetter, Val); 1676 } 1677 } 1678 CSInfo.markDevirt(); 1679 } 1680 1681 bool DevirtModule::tryVirtualConstProp( 1682 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1683 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1684 // This only works if the function returns an integer. 1685 auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType()); 1686 if (!RetType) 1687 return false; 1688 unsigned BitWidth = RetType->getBitWidth(); 1689 if (BitWidth > 64) 1690 return false; 1691 1692 // Make sure that each function is defined, does not access memory, takes at 1693 // least one argument, does not use its first argument (which we assume is 1694 // 'this'), and has the same return type. 1695 // 1696 // Note that we test whether this copy of the function is readnone, rather 1697 // than testing function attributes, which must hold for any copy of the 1698 // function, even a less optimized version substituted at link time. This is 1699 // sound because the virtual constant propagation optimizations effectively 1700 // inline all implementations of the virtual function into each call site, 1701 // rather than using function attributes to perform local optimization. 1702 for (VirtualCallTarget &Target : TargetsForSlot) { 1703 if (Target.Fn->isDeclaration() || 1704 computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) != 1705 MAK_ReadNone || 1706 Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() || 1707 Target.Fn->getReturnType() != RetType) 1708 return false; 1709 } 1710 1711 for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) { 1712 if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first)) 1713 continue; 1714 1715 WholeProgramDevirtResolution::ByArg *ResByArg = nullptr; 1716 if (Res) 1717 ResByArg = &Res->ResByArg[CSByConstantArg.first]; 1718 1719 if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg)) 1720 continue; 1721 1722 if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second, 1723 ResByArg, Slot, CSByConstantArg.first)) 1724 continue; 1725 1726 // Find an allocation offset in bits in all vtables associated with the 1727 // type. 1728 uint64_t AllocBefore = 1729 findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth); 1730 uint64_t AllocAfter = 1731 findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth); 1732 1733 // Calculate the total amount of padding needed to store a value at both 1734 // ends of the object. 1735 uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0; 1736 for (auto &&Target : TargetsForSlot) { 1737 TotalPaddingBefore += std::max<int64_t>( 1738 (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0); 1739 TotalPaddingAfter += std::max<int64_t>( 1740 (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0); 1741 } 1742 1743 // If the amount of padding is too large, give up. 1744 // FIXME: do something smarter here. 1745 if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128) 1746 continue; 1747 1748 // Calculate the offset to the value as a (possibly negative) byte offset 1749 // and (if applicable) a bit offset, and store the values in the targets. 1750 int64_t OffsetByte; 1751 uint64_t OffsetBit; 1752 if (TotalPaddingBefore <= TotalPaddingAfter) 1753 setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte, 1754 OffsetBit); 1755 else 1756 setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte, 1757 OffsetBit); 1758 1759 if (RemarksEnabled) 1760 for (auto &&Target : TargetsForSlot) 1761 Target.WasDevirt = true; 1762 1763 1764 if (CSByConstantArg.second.isExported()) { 1765 ResByArg->TheKind = WholeProgramDevirtResolution::ByArg::VirtualConstProp; 1766 exportConstant(Slot, CSByConstantArg.first, "byte", OffsetByte, 1767 ResByArg->Byte); 1768 exportConstant(Slot, CSByConstantArg.first, "bit", 1ULL << OffsetBit, 1769 ResByArg->Bit); 1770 } 1771 1772 // Rewrite each call to a load from OffsetByte/OffsetBit. 1773 Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte); 1774 Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit); 1775 applyVirtualConstProp(CSByConstantArg.second, 1776 TargetsForSlot[0].Fn->getName(), ByteConst, BitConst); 1777 } 1778 return true; 1779 } 1780 1781 void DevirtModule::rebuildGlobal(VTableBits &B) { 1782 if (B.Before.Bytes.empty() && B.After.Bytes.empty()) 1783 return; 1784 1785 // Align the before byte array to the global's minimum alignment so that we 1786 // don't break any alignment requirements on the global. 1787 Align Alignment = M.getDataLayout().getValueOrABITypeAlignment( 1788 B.GV->getAlign(), B.GV->getValueType()); 1789 B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), Alignment)); 1790 1791 // Before was stored in reverse order; flip it now. 1792 for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I) 1793 std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]); 1794 1795 // Build an anonymous global containing the before bytes, followed by the 1796 // original initializer, followed by the after bytes. 1797 auto NewInit = ConstantStruct::getAnon( 1798 {ConstantDataArray::get(M.getContext(), B.Before.Bytes), 1799 B.GV->getInitializer(), 1800 ConstantDataArray::get(M.getContext(), B.After.Bytes)}); 1801 auto NewGV = 1802 new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(), 1803 GlobalVariable::PrivateLinkage, NewInit, "", B.GV); 1804 NewGV->setSection(B.GV->getSection()); 1805 NewGV->setComdat(B.GV->getComdat()); 1806 NewGV->setAlignment(B.GV->getAlign()); 1807 1808 // Copy the original vtable's metadata to the anonymous global, adjusting 1809 // offsets as required. 1810 NewGV->copyMetadata(B.GV, B.Before.Bytes.size()); 1811 1812 // Build an alias named after the original global, pointing at the second 1813 // element (the original initializer). 1814 auto Alias = GlobalAlias::create( 1815 B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "", 1816 ConstantExpr::getGetElementPtr( 1817 NewInit->getType(), NewGV, 1818 ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0), 1819 ConstantInt::get(Int32Ty, 1)}), 1820 &M); 1821 Alias->setVisibility(B.GV->getVisibility()); 1822 Alias->takeName(B.GV); 1823 1824 B.GV->replaceAllUsesWith(Alias); 1825 B.GV->eraseFromParent(); 1826 } 1827 1828 bool DevirtModule::areRemarksEnabled() { 1829 const auto &FL = M.getFunctionList(); 1830 for (const Function &Fn : FL) { 1831 const auto &BBL = Fn.getBasicBlockList(); 1832 if (BBL.empty()) 1833 continue; 1834 auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front()); 1835 return DI.isEnabled(); 1836 } 1837 return false; 1838 } 1839 1840 void DevirtModule::scanTypeTestUsers( 1841 Function *TypeTestFunc, 1842 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 1843 // Find all virtual calls via a virtual table pointer %p under an assumption 1844 // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p 1845 // points to a member of the type identifier %md. Group calls by (type ID, 1846 // offset) pair (effectively the identity of the virtual function) and store 1847 // to CallSlots. 1848 for (Use &U : llvm::make_early_inc_range(TypeTestFunc->uses())) { 1849 auto *CI = dyn_cast<CallInst>(U.getUser()); 1850 if (!CI) 1851 continue; 1852 1853 // Search for virtual calls based on %p and add them to DevirtCalls. 1854 SmallVector<DevirtCallSite, 1> DevirtCalls; 1855 SmallVector<CallInst *, 1> Assumes; 1856 auto &DT = LookupDomTree(*CI->getFunction()); 1857 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT); 1858 1859 Metadata *TypeId = 1860 cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata(); 1861 // If we found any, add them to CallSlots. 1862 if (!Assumes.empty()) { 1863 Value *Ptr = CI->getArgOperand(0)->stripPointerCasts(); 1864 for (DevirtCallSite Call : DevirtCalls) 1865 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, nullptr); 1866 } 1867 1868 auto RemoveTypeTestAssumes = [&]() { 1869 // We no longer need the assumes or the type test. 1870 for (auto Assume : Assumes) 1871 Assume->eraseFromParent(); 1872 // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we 1873 // may use the vtable argument later. 1874 if (CI->use_empty()) 1875 CI->eraseFromParent(); 1876 }; 1877 1878 // At this point we could remove all type test assume sequences, as they 1879 // were originally inserted for WPD. However, we can keep these in the 1880 // code stream for later analysis (e.g. to help drive more efficient ICP 1881 // sequences). They will eventually be removed by a second LowerTypeTests 1882 // invocation that cleans them up. In order to do this correctly, the first 1883 // LowerTypeTests invocation needs to know that they have "Unknown" type 1884 // test resolution, so that they aren't treated as Unsat and lowered to 1885 // False, which will break any uses on assumes. Below we remove any type 1886 // test assumes that will not be treated as Unknown by LTT. 1887 1888 // The type test assumes will be treated by LTT as Unsat if the type id is 1889 // not used on a global (in which case it has no entry in the TypeIdMap). 1890 if (!TypeIdMap.count(TypeId)) 1891 RemoveTypeTestAssumes(); 1892 1893 // For ThinLTO importing, we need to remove the type test assumes if this is 1894 // an MDString type id without a corresponding TypeIdSummary. Any 1895 // non-MDString type ids are ignored and treated as Unknown by LTT, so their 1896 // type test assumes can be kept. If the MDString type id is missing a 1897 // TypeIdSummary (e.g. because there was no use on a vcall, preventing the 1898 // exporting phase of WPD from analyzing it), then it would be treated as 1899 // Unsat by LTT and we need to remove its type test assumes here. If not 1900 // used on a vcall we don't need them for later optimization use in any 1901 // case. 1902 else if (ImportSummary && isa<MDString>(TypeId)) { 1903 const TypeIdSummary *TidSummary = 1904 ImportSummary->getTypeIdSummary(cast<MDString>(TypeId)->getString()); 1905 if (!TidSummary) 1906 RemoveTypeTestAssumes(); 1907 else 1908 // If one was created it should not be Unsat, because if we reached here 1909 // the type id was used on a global. 1910 assert(TidSummary->TTRes.TheKind != TypeTestResolution::Unsat); 1911 } 1912 } 1913 } 1914 1915 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) { 1916 Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test); 1917 1918 for (Use &U : llvm::make_early_inc_range(TypeCheckedLoadFunc->uses())) { 1919 auto *CI = dyn_cast<CallInst>(U.getUser()); 1920 if (!CI) 1921 continue; 1922 1923 Value *Ptr = CI->getArgOperand(0); 1924 Value *Offset = CI->getArgOperand(1); 1925 Value *TypeIdValue = CI->getArgOperand(2); 1926 Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata(); 1927 1928 SmallVector<DevirtCallSite, 1> DevirtCalls; 1929 SmallVector<Instruction *, 1> LoadedPtrs; 1930 SmallVector<Instruction *, 1> Preds; 1931 bool HasNonCallUses = false; 1932 auto &DT = LookupDomTree(*CI->getFunction()); 1933 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds, 1934 HasNonCallUses, CI, DT); 1935 1936 // Start by generating "pessimistic" code that explicitly loads the function 1937 // pointer from the vtable and performs the type check. If possible, we will 1938 // eliminate the load and the type check later. 1939 1940 // If possible, only generate the load at the point where it is used. 1941 // This helps avoid unnecessary spills. 1942 IRBuilder<> LoadB( 1943 (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI); 1944 Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset); 1945 Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy)); 1946 Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr); 1947 1948 for (Instruction *LoadedPtr : LoadedPtrs) { 1949 LoadedPtr->replaceAllUsesWith(LoadedValue); 1950 LoadedPtr->eraseFromParent(); 1951 } 1952 1953 // Likewise for the type test. 1954 IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI); 1955 CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue}); 1956 1957 for (Instruction *Pred : Preds) { 1958 Pred->replaceAllUsesWith(TypeTestCall); 1959 Pred->eraseFromParent(); 1960 } 1961 1962 // We have already erased any extractvalue instructions that refer to the 1963 // intrinsic call, but the intrinsic may have other non-extractvalue uses 1964 // (although this is unlikely). In that case, explicitly build a pair and 1965 // RAUW it. 1966 if (!CI->use_empty()) { 1967 Value *Pair = UndefValue::get(CI->getType()); 1968 IRBuilder<> B(CI); 1969 Pair = B.CreateInsertValue(Pair, LoadedValue, {0}); 1970 Pair = B.CreateInsertValue(Pair, TypeTestCall, {1}); 1971 CI->replaceAllUsesWith(Pair); 1972 } 1973 1974 // The number of unsafe uses is initially the number of uses. 1975 auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall]; 1976 NumUnsafeUses = DevirtCalls.size(); 1977 1978 // If the function pointer has a non-call user, we cannot eliminate the type 1979 // check, as one of those users may eventually call the pointer. Increment 1980 // the unsafe use count to make sure it cannot reach zero. 1981 if (HasNonCallUses) 1982 ++NumUnsafeUses; 1983 for (DevirtCallSite Call : DevirtCalls) { 1984 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, 1985 &NumUnsafeUses); 1986 } 1987 1988 CI->eraseFromParent(); 1989 } 1990 } 1991 1992 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) { 1993 auto *TypeId = dyn_cast<MDString>(Slot.TypeID); 1994 if (!TypeId) 1995 return; 1996 const TypeIdSummary *TidSummary = 1997 ImportSummary->getTypeIdSummary(TypeId->getString()); 1998 if (!TidSummary) 1999 return; 2000 auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset); 2001 if (ResI == TidSummary->WPDRes.end()) 2002 return; 2003 const WholeProgramDevirtResolution &Res = ResI->second; 2004 2005 if (Res.TheKind == WholeProgramDevirtResolution::SingleImpl) { 2006 assert(!Res.SingleImplName.empty()); 2007 // The type of the function in the declaration is irrelevant because every 2008 // call site will cast it to the correct type. 2009 Constant *SingleImpl = 2010 cast<Constant>(M.getOrInsertFunction(Res.SingleImplName, 2011 Type::getVoidTy(M.getContext())) 2012 .getCallee()); 2013 2014 // This is the import phase so we should not be exporting anything. 2015 bool IsExported = false; 2016 applySingleImplDevirt(SlotInfo, SingleImpl, IsExported); 2017 assert(!IsExported); 2018 } 2019 2020 for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) { 2021 auto I = Res.ResByArg.find(CSByConstantArg.first); 2022 if (I == Res.ResByArg.end()) 2023 continue; 2024 auto &ResByArg = I->second; 2025 // FIXME: We should figure out what to do about the "function name" argument 2026 // to the apply* functions, as the function names are unavailable during the 2027 // importing phase. For now we just pass the empty string. This does not 2028 // impact correctness because the function names are just used for remarks. 2029 switch (ResByArg.TheKind) { 2030 case WholeProgramDevirtResolution::ByArg::UniformRetVal: 2031 applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info); 2032 break; 2033 case WholeProgramDevirtResolution::ByArg::UniqueRetVal: { 2034 Constant *UniqueMemberAddr = 2035 importGlobal(Slot, CSByConstantArg.first, "unique_member"); 2036 applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info, 2037 UniqueMemberAddr); 2038 break; 2039 } 2040 case WholeProgramDevirtResolution::ByArg::VirtualConstProp: { 2041 Constant *Byte = importConstant(Slot, CSByConstantArg.first, "byte", 2042 Int32Ty, ResByArg.Byte); 2043 Constant *Bit = importConstant(Slot, CSByConstantArg.first, "bit", Int8Ty, 2044 ResByArg.Bit); 2045 applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit); 2046 break; 2047 } 2048 default: 2049 break; 2050 } 2051 } 2052 2053 if (Res.TheKind == WholeProgramDevirtResolution::BranchFunnel) { 2054 // The type of the function is irrelevant, because it's bitcast at calls 2055 // anyhow. 2056 Constant *JT = cast<Constant>( 2057 M.getOrInsertFunction(getGlobalName(Slot, {}, "branch_funnel"), 2058 Type::getVoidTy(M.getContext())) 2059 .getCallee()); 2060 bool IsExported = false; 2061 applyICallBranchFunnel(SlotInfo, JT, IsExported); 2062 assert(!IsExported); 2063 } 2064 } 2065 2066 void DevirtModule::removeRedundantTypeTests() { 2067 auto True = ConstantInt::getTrue(M.getContext()); 2068 for (auto &&U : NumUnsafeUsesForTypeTest) { 2069 if (U.second == 0) { 2070 U.first->replaceAllUsesWith(True); 2071 U.first->eraseFromParent(); 2072 } 2073 } 2074 } 2075 2076 ValueInfo 2077 DevirtModule::lookUpFunctionValueInfo(Function *TheFn, 2078 ModuleSummaryIndex *ExportSummary) { 2079 assert((ExportSummary != nullptr) && 2080 "Caller guarantees ExportSummary is not nullptr"); 2081 2082 const auto TheFnGUID = TheFn->getGUID(); 2083 const auto TheFnGUIDWithExportedName = GlobalValue::getGUID(TheFn->getName()); 2084 // Look up ValueInfo with the GUID in the current linkage. 2085 ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFnGUID); 2086 // If no entry is found and GUID is different from GUID computed using 2087 // exported name, look up ValueInfo with the exported name unconditionally. 2088 // This is a fallback. 2089 // 2090 // The reason to have a fallback: 2091 // 1. LTO could enable global value internalization via 2092 // `enable-lto-internalization`. 2093 // 2. The GUID in ExportedSummary is computed using exported name. 2094 if ((!TheFnVI) && (TheFnGUID != TheFnGUIDWithExportedName)) { 2095 TheFnVI = ExportSummary->getValueInfo(TheFnGUIDWithExportedName); 2096 } 2097 return TheFnVI; 2098 } 2099 2100 bool DevirtModule::mustBeUnreachableFunction( 2101 Function *const F, ModuleSummaryIndex *ExportSummary) { 2102 // First, learn unreachability by analyzing function IR. 2103 if (!F->isDeclaration()) { 2104 // A function must be unreachable if its entry block ends with an 2105 // 'unreachable'. 2106 return isa<UnreachableInst>(F->getEntryBlock().getTerminator()); 2107 } 2108 // Learn unreachability from ExportSummary if ExportSummary is present. 2109 return ExportSummary && 2110 ::mustBeUnreachableFunction( 2111 DevirtModule::lookUpFunctionValueInfo(F, ExportSummary)); 2112 } 2113 2114 bool DevirtModule::run() { 2115 // If only some of the modules were split, we cannot correctly perform 2116 // this transformation. We already checked for the presense of type tests 2117 // with partially split modules during the thin link, and would have emitted 2118 // an error if any were found, so here we can simply return. 2119 if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) || 2120 (ImportSummary && ImportSummary->partiallySplitLTOUnits())) 2121 return false; 2122 2123 Function *TypeTestFunc = 2124 M.getFunction(Intrinsic::getName(Intrinsic::type_test)); 2125 Function *TypeCheckedLoadFunc = 2126 M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load)); 2127 Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume)); 2128 2129 // Normally if there are no users of the devirtualization intrinsics in the 2130 // module, this pass has nothing to do. But if we are exporting, we also need 2131 // to handle any users that appear only in the function summaries. 2132 if (!ExportSummary && 2133 (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc || 2134 AssumeFunc->use_empty()) && 2135 (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty())) 2136 return false; 2137 2138 // Rebuild type metadata into a map for easy lookup. 2139 std::vector<VTableBits> Bits; 2140 DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap; 2141 buildTypeIdentifierMap(Bits, TypeIdMap); 2142 2143 if (TypeTestFunc && AssumeFunc) 2144 scanTypeTestUsers(TypeTestFunc, TypeIdMap); 2145 2146 if (TypeCheckedLoadFunc) 2147 scanTypeCheckedLoadUsers(TypeCheckedLoadFunc); 2148 2149 if (ImportSummary) { 2150 for (auto &S : CallSlots) 2151 importResolution(S.first, S.second); 2152 2153 removeRedundantTypeTests(); 2154 2155 // We have lowered or deleted the type instrinsics, so we will no 2156 // longer have enough information to reason about the liveness of virtual 2157 // function pointers in GlobalDCE. 2158 for (GlobalVariable &GV : M.globals()) 2159 GV.eraseMetadata(LLVMContext::MD_vcall_visibility); 2160 2161 // The rest of the code is only necessary when exporting or during regular 2162 // LTO, so we are done. 2163 return true; 2164 } 2165 2166 if (TypeIdMap.empty()) 2167 return true; 2168 2169 // Collect information from summary about which calls to try to devirtualize. 2170 if (ExportSummary) { 2171 DenseMap<GlobalValue::GUID, TinyPtrVector<Metadata *>> MetadataByGUID; 2172 for (auto &P : TypeIdMap) { 2173 if (auto *TypeId = dyn_cast<MDString>(P.first)) 2174 MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back( 2175 TypeId); 2176 } 2177 2178 for (auto &P : *ExportSummary) { 2179 for (auto &S : P.second.SummaryList) { 2180 auto *FS = dyn_cast<FunctionSummary>(S.get()); 2181 if (!FS) 2182 continue; 2183 // FIXME: Only add live functions. 2184 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 2185 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 2186 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 2187 } 2188 } 2189 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 2190 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 2191 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 2192 } 2193 } 2194 for (const FunctionSummary::ConstVCall &VC : 2195 FS->type_test_assume_const_vcalls()) { 2196 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 2197 CallSlots[{MD, VC.VFunc.Offset}] 2198 .ConstCSInfo[VC.Args] 2199 .addSummaryTypeTestAssumeUser(FS); 2200 } 2201 } 2202 for (const FunctionSummary::ConstVCall &VC : 2203 FS->type_checked_load_const_vcalls()) { 2204 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 2205 CallSlots[{MD, VC.VFunc.Offset}] 2206 .ConstCSInfo[VC.Args] 2207 .addSummaryTypeCheckedLoadUser(FS); 2208 } 2209 } 2210 } 2211 } 2212 } 2213 2214 // For each (type, offset) pair: 2215 bool DidVirtualConstProp = false; 2216 std::map<std::string, Function*> DevirtTargets; 2217 for (auto &S : CallSlots) { 2218 // Search each of the members of the type identifier for the virtual 2219 // function implementation at offset S.first.ByteOffset, and add to 2220 // TargetsForSlot. 2221 std::vector<VirtualCallTarget> TargetsForSlot; 2222 WholeProgramDevirtResolution *Res = nullptr; 2223 const std::set<TypeMemberInfo> &TypeMemberInfos = TypeIdMap[S.first.TypeID]; 2224 if (ExportSummary && isa<MDString>(S.first.TypeID) && 2225 TypeMemberInfos.size()) 2226 // For any type id used on a global's type metadata, create the type id 2227 // summary resolution regardless of whether we can devirtualize, so that 2228 // lower type tests knows the type id is not Unsat. If it was not used on 2229 // a global's type metadata, the TypeIdMap entry set will be empty, and 2230 // we don't want to create an entry (with the default Unknown type 2231 // resolution), which can prevent detection of the Unsat. 2232 Res = &ExportSummary 2233 ->getOrInsertTypeIdSummary( 2234 cast<MDString>(S.first.TypeID)->getString()) 2235 .WPDRes[S.first.ByteOffset]; 2236 if (tryFindVirtualCallTargets(TargetsForSlot, TypeMemberInfos, 2237 S.first.ByteOffset, ExportSummary)) { 2238 2239 if (!trySingleImplDevirt(ExportSummary, TargetsForSlot, S.second, Res)) { 2240 DidVirtualConstProp |= 2241 tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first); 2242 2243 tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first); 2244 } 2245 2246 // Collect functions devirtualized at least for one call site for stats. 2247 if (RemarksEnabled) 2248 for (const auto &T : TargetsForSlot) 2249 if (T.WasDevirt) 2250 DevirtTargets[std::string(T.Fn->getName())] = T.Fn; 2251 } 2252 2253 // CFI-specific: if we are exporting and any llvm.type.checked.load 2254 // intrinsics were *not* devirtualized, we need to add the resulting 2255 // llvm.type.test intrinsics to the function summaries so that the 2256 // LowerTypeTests pass will export them. 2257 if (ExportSummary && isa<MDString>(S.first.TypeID)) { 2258 auto GUID = 2259 GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString()); 2260 for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers) 2261 FS->addTypeTest(GUID); 2262 for (auto &CCS : S.second.ConstCSInfo) 2263 for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers) 2264 FS->addTypeTest(GUID); 2265 } 2266 } 2267 2268 if (RemarksEnabled) { 2269 // Generate remarks for each devirtualized function. 2270 for (const auto &DT : DevirtTargets) { 2271 Function *F = DT.second; 2272 2273 using namespace ore; 2274 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F) 2275 << "devirtualized " 2276 << NV("FunctionName", DT.first)); 2277 } 2278 } 2279 2280 removeRedundantTypeTests(); 2281 2282 // Rebuild each global we touched as part of virtual constant propagation to 2283 // include the before and after bytes. 2284 if (DidVirtualConstProp) 2285 for (VTableBits &B : Bits) 2286 rebuildGlobal(B); 2287 2288 // We have lowered or deleted the type instrinsics, so we will no 2289 // longer have enough information to reason about the liveness of virtual 2290 // function pointers in GlobalDCE. 2291 for (GlobalVariable &GV : M.globals()) 2292 GV.eraseMetadata(LLVMContext::MD_vcall_visibility); 2293 2294 return true; 2295 } 2296 2297 void DevirtIndex::run() { 2298 if (ExportSummary.typeIdCompatibleVtableMap().empty()) 2299 return; 2300 2301 DenseMap<GlobalValue::GUID, std::vector<StringRef>> NameByGUID; 2302 for (auto &P : ExportSummary.typeIdCompatibleVtableMap()) { 2303 NameByGUID[GlobalValue::getGUID(P.first)].push_back(P.first); 2304 } 2305 2306 // Collect information from summary about which calls to try to devirtualize. 2307 for (auto &P : ExportSummary) { 2308 for (auto &S : P.second.SummaryList) { 2309 auto *FS = dyn_cast<FunctionSummary>(S.get()); 2310 if (!FS) 2311 continue; 2312 // FIXME: Only add live functions. 2313 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 2314 for (StringRef Name : NameByGUID[VF.GUID]) { 2315 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 2316 } 2317 } 2318 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 2319 for (StringRef Name : NameByGUID[VF.GUID]) { 2320 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 2321 } 2322 } 2323 for (const FunctionSummary::ConstVCall &VC : 2324 FS->type_test_assume_const_vcalls()) { 2325 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 2326 CallSlots[{Name, VC.VFunc.Offset}] 2327 .ConstCSInfo[VC.Args] 2328 .addSummaryTypeTestAssumeUser(FS); 2329 } 2330 } 2331 for (const FunctionSummary::ConstVCall &VC : 2332 FS->type_checked_load_const_vcalls()) { 2333 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 2334 CallSlots[{Name, VC.VFunc.Offset}] 2335 .ConstCSInfo[VC.Args] 2336 .addSummaryTypeCheckedLoadUser(FS); 2337 } 2338 } 2339 } 2340 } 2341 2342 std::set<ValueInfo> DevirtTargets; 2343 // For each (type, offset) pair: 2344 for (auto &S : CallSlots) { 2345 // Search each of the members of the type identifier for the virtual 2346 // function implementation at offset S.first.ByteOffset, and add to 2347 // TargetsForSlot. 2348 std::vector<ValueInfo> TargetsForSlot; 2349 auto TidSummary = ExportSummary.getTypeIdCompatibleVtableSummary(S.first.TypeID); 2350 assert(TidSummary); 2351 // Create the type id summary resolution regardlness of whether we can 2352 // devirtualize, so that lower type tests knows the type id is used on 2353 // a global and not Unsat. 2354 WholeProgramDevirtResolution *Res = 2355 &ExportSummary.getOrInsertTypeIdSummary(S.first.TypeID) 2356 .WPDRes[S.first.ByteOffset]; 2357 if (tryFindVirtualCallTargets(TargetsForSlot, *TidSummary, 2358 S.first.ByteOffset)) { 2359 2360 if (!trySingleImplDevirt(TargetsForSlot, S.first, S.second, Res, 2361 DevirtTargets)) 2362 continue; 2363 } 2364 } 2365 2366 // Optionally have the thin link print message for each devirtualized 2367 // function. 2368 if (PrintSummaryDevirt) 2369 for (const auto &DT : DevirtTargets) 2370 errs() << "Devirtualized call to " << DT << "\n"; 2371 } 2372