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/Statistic.h" 61 #include "llvm/ADT/Triple.h" 62 #include "llvm/ADT/iterator_range.h" 63 #include "llvm/Analysis/AssumptionCache.h" 64 #include "llvm/Analysis/BasicAliasAnalysis.h" 65 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 66 #include "llvm/Analysis/TypeMetadataUtils.h" 67 #include "llvm/Bitcode/BitcodeReader.h" 68 #include "llvm/Bitcode/BitcodeWriter.h" 69 #include "llvm/IR/Constants.h" 70 #include "llvm/IR/DataLayout.h" 71 #include "llvm/IR/DebugLoc.h" 72 #include "llvm/IR/DerivedTypes.h" 73 #include "llvm/IR/Dominators.h" 74 #include "llvm/IR/Function.h" 75 #include "llvm/IR/GlobalAlias.h" 76 #include "llvm/IR/GlobalVariable.h" 77 #include "llvm/IR/IRBuilder.h" 78 #include "llvm/IR/InstrTypes.h" 79 #include "llvm/IR/Instruction.h" 80 #include "llvm/IR/Instructions.h" 81 #include "llvm/IR/Intrinsics.h" 82 #include "llvm/IR/LLVMContext.h" 83 #include "llvm/IR/MDBuilder.h" 84 #include "llvm/IR/Metadata.h" 85 #include "llvm/IR/Module.h" 86 #include "llvm/IR/ModuleSummaryIndexYAML.h" 87 #include "llvm/InitializePasses.h" 88 #include "llvm/Pass.h" 89 #include "llvm/PassRegistry.h" 90 #include "llvm/Support/Casting.h" 91 #include "llvm/Support/CommandLine.h" 92 #include "llvm/Support/Errc.h" 93 #include "llvm/Support/Error.h" 94 #include "llvm/Support/FileSystem.h" 95 #include "llvm/Support/GlobPattern.h" 96 #include "llvm/Support/MathExtras.h" 97 #include "llvm/Transforms/IPO.h" 98 #include "llvm/Transforms/IPO/FunctionAttrs.h" 99 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 100 #include "llvm/Transforms/Utils/CallPromotionUtils.h" 101 #include "llvm/Transforms/Utils/Evaluator.h" 102 #include <algorithm> 103 #include <cstddef> 104 #include <map> 105 #include <set> 106 #include <string> 107 108 using namespace llvm; 109 using namespace wholeprogramdevirt; 110 111 #define DEBUG_TYPE "wholeprogramdevirt" 112 113 STATISTIC(NumDevirtTargets, "Number of whole program devirtualization targets"); 114 STATISTIC(NumSingleImpl, "Number of single implementation devirtualizations"); 115 STATISTIC(NumBranchFunnel, "Number of branch funnels"); 116 STATISTIC(NumUniformRetVal, "Number of uniform return value optimizations"); 117 STATISTIC(NumUniqueRetVal, "Number of unique return value optimizations"); 118 STATISTIC(NumVirtConstProp1Bit, 119 "Number of 1 bit virtual constant propagations"); 120 STATISTIC(NumVirtConstProp, "Number of virtual constant propagations"); 121 122 static cl::opt<PassSummaryAction> ClSummaryAction( 123 "wholeprogramdevirt-summary-action", 124 cl::desc("What to do with the summary when running this pass"), 125 cl::values(clEnumValN(PassSummaryAction::None, "none", "Do nothing"), 126 clEnumValN(PassSummaryAction::Import, "import", 127 "Import typeid resolutions from summary and globals"), 128 clEnumValN(PassSummaryAction::Export, "export", 129 "Export typeid resolutions to summary and globals")), 130 cl::Hidden); 131 132 static cl::opt<std::string> ClReadSummary( 133 "wholeprogramdevirt-read-summary", 134 cl::desc( 135 "Read summary from given bitcode or YAML file before running pass"), 136 cl::Hidden); 137 138 static cl::opt<std::string> ClWriteSummary( 139 "wholeprogramdevirt-write-summary", 140 cl::desc("Write summary to given bitcode or YAML file after running pass. " 141 "Output file format is deduced from extension: *.bc means writing " 142 "bitcode, otherwise YAML"), 143 cl::Hidden); 144 145 static cl::opt<unsigned> 146 ClThreshold("wholeprogramdevirt-branch-funnel-threshold", cl::Hidden, 147 cl::init(10), 148 cl::desc("Maximum number of call targets per " 149 "call site to enable branch funnels")); 150 151 static cl::opt<bool> 152 PrintSummaryDevirt("wholeprogramdevirt-print-index-based", cl::Hidden, 153 cl::desc("Print index-based devirtualization messages")); 154 155 /// Provide a way to force enable whole program visibility in tests. 156 /// This is needed to support legacy tests that don't contain 157 /// !vcall_visibility metadata (the mere presense of type tests 158 /// previously implied hidden visibility). 159 static cl::opt<bool> 160 WholeProgramVisibility("whole-program-visibility", cl::Hidden, 161 cl::desc("Enable whole program visibility")); 162 163 /// Provide a way to force disable whole program for debugging or workarounds, 164 /// when enabled via the linker. 165 static cl::opt<bool> DisableWholeProgramVisibility( 166 "disable-whole-program-visibility", cl::Hidden, 167 cl::desc("Disable whole program visibility (overrides enabling options)")); 168 169 /// Provide way to prevent certain function from being devirtualized 170 static cl::list<std::string> 171 SkipFunctionNames("wholeprogramdevirt-skip", 172 cl::desc("Prevent function(s) from being devirtualized"), 173 cl::Hidden, cl::CommaSeparated); 174 175 /// Mechanism to add runtime checking of devirtualization decisions, optionally 176 /// trapping or falling back to indirect call on any that are not correct. 177 /// Trapping mode is useful for debugging undefined behavior leading to failures 178 /// with WPD. Fallback mode is useful for ensuring safety when whole program 179 /// visibility may be compromised. 180 enum WPDCheckMode { None, Trap, Fallback }; 181 static cl::opt<WPDCheckMode> DevirtCheckMode( 182 "wholeprogramdevirt-check", cl::Hidden, 183 cl::desc("Type of checking for incorrect devirtualizations"), 184 cl::values(clEnumValN(WPDCheckMode::None, "none", "No checking"), 185 clEnumValN(WPDCheckMode::Trap, "trap", "Trap when incorrect"), 186 clEnumValN(WPDCheckMode::Fallback, "fallback", 187 "Fallback to indirect when incorrect"))); 188 189 namespace { 190 struct PatternList { 191 std::vector<GlobPattern> Patterns; 192 template <class T> void init(const T &StringList) { 193 for (const auto &S : StringList) 194 if (Expected<GlobPattern> Pat = GlobPattern::create(S)) 195 Patterns.push_back(std::move(*Pat)); 196 } 197 bool match(StringRef S) { 198 for (const GlobPattern &P : Patterns) 199 if (P.match(S)) 200 return true; 201 return false; 202 } 203 }; 204 } // namespace 205 206 // Find the minimum offset that we may store a value of size Size bits at. If 207 // IsAfter is set, look for an offset before the object, otherwise look for an 208 // offset after the object. 209 uint64_t 210 wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets, 211 bool IsAfter, uint64_t Size) { 212 // Find a minimum offset taking into account only vtable sizes. 213 uint64_t MinByte = 0; 214 for (const VirtualCallTarget &Target : Targets) { 215 if (IsAfter) 216 MinByte = std::max(MinByte, Target.minAfterBytes()); 217 else 218 MinByte = std::max(MinByte, Target.minBeforeBytes()); 219 } 220 221 // Build a vector of arrays of bytes covering, for each target, a slice of the 222 // used region (see AccumBitVector::BytesUsed in 223 // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively, 224 // this aligns the used regions to start at MinByte. 225 // 226 // In this example, A, B and C are vtables, # is a byte already allocated for 227 // a virtual function pointer, AAAA... (etc.) are the used regions for the 228 // vtables and Offset(X) is the value computed for the Offset variable below 229 // for X. 230 // 231 // Offset(A) 232 // | | 233 // |MinByte 234 // A: ################AAAAAAAA|AAAAAAAA 235 // B: ########BBBBBBBBBBBBBBBB|BBBB 236 // C: ########################|CCCCCCCCCCCCCCCC 237 // | Offset(B) | 238 // 239 // This code produces the slices of A, B and C that appear after the divider 240 // at MinByte. 241 std::vector<ArrayRef<uint8_t>> Used; 242 for (const VirtualCallTarget &Target : Targets) { 243 ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed 244 : Target.TM->Bits->Before.BytesUsed; 245 uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes() 246 : MinByte - Target.minBeforeBytes(); 247 248 // Disregard used regions that are smaller than Offset. These are 249 // effectively all-free regions that do not need to be checked. 250 if (VTUsed.size() > Offset) 251 Used.push_back(VTUsed.slice(Offset)); 252 } 253 254 if (Size == 1) { 255 // Find a free bit in each member of Used. 256 for (unsigned I = 0;; ++I) { 257 uint8_t BitsUsed = 0; 258 for (auto &&B : Used) 259 if (I < B.size()) 260 BitsUsed |= B[I]; 261 if (BitsUsed != 0xff) 262 return (MinByte + I) * 8 + 263 countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined); 264 } 265 } else { 266 // Find a free (Size/8) byte region in each member of Used. 267 // FIXME: see if alignment helps. 268 for (unsigned I = 0;; ++I) { 269 for (auto &&B : Used) { 270 unsigned Byte = 0; 271 while ((I + Byte) < B.size() && Byte < (Size / 8)) { 272 if (B[I + Byte]) 273 goto NextI; 274 ++Byte; 275 } 276 } 277 return (MinByte + I) * 8; 278 NextI:; 279 } 280 } 281 } 282 283 void wholeprogramdevirt::setBeforeReturnValues( 284 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore, 285 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 286 if (BitWidth == 1) 287 OffsetByte = -(AllocBefore / 8 + 1); 288 else 289 OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8); 290 OffsetBit = AllocBefore % 8; 291 292 for (VirtualCallTarget &Target : Targets) { 293 if (BitWidth == 1) 294 Target.setBeforeBit(AllocBefore); 295 else 296 Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8); 297 } 298 } 299 300 void wholeprogramdevirt::setAfterReturnValues( 301 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter, 302 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 303 if (BitWidth == 1) 304 OffsetByte = AllocAfter / 8; 305 else 306 OffsetByte = (AllocAfter + 7) / 8; 307 OffsetBit = AllocAfter % 8; 308 309 for (VirtualCallTarget &Target : Targets) { 310 if (BitWidth == 1) 311 Target.setAfterBit(AllocAfter); 312 else 313 Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8); 314 } 315 } 316 317 VirtualCallTarget::VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM) 318 : Fn(Fn), TM(TM), 319 IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {} 320 321 namespace { 322 323 // A slot in a set of virtual tables. The TypeID identifies the set of virtual 324 // tables, and the ByteOffset is the offset in bytes from the address point to 325 // the virtual function pointer. 326 struct VTableSlot { 327 Metadata *TypeID; 328 uint64_t ByteOffset; 329 }; 330 331 } // end anonymous namespace 332 333 namespace llvm { 334 335 template <> struct DenseMapInfo<VTableSlot> { 336 static VTableSlot getEmptyKey() { 337 return {DenseMapInfo<Metadata *>::getEmptyKey(), 338 DenseMapInfo<uint64_t>::getEmptyKey()}; 339 } 340 static VTableSlot getTombstoneKey() { 341 return {DenseMapInfo<Metadata *>::getTombstoneKey(), 342 DenseMapInfo<uint64_t>::getTombstoneKey()}; 343 } 344 static unsigned getHashValue(const VTableSlot &I) { 345 return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^ 346 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 347 } 348 static bool isEqual(const VTableSlot &LHS, 349 const VTableSlot &RHS) { 350 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 351 } 352 }; 353 354 template <> struct DenseMapInfo<VTableSlotSummary> { 355 static VTableSlotSummary getEmptyKey() { 356 return {DenseMapInfo<StringRef>::getEmptyKey(), 357 DenseMapInfo<uint64_t>::getEmptyKey()}; 358 } 359 static VTableSlotSummary getTombstoneKey() { 360 return {DenseMapInfo<StringRef>::getTombstoneKey(), 361 DenseMapInfo<uint64_t>::getTombstoneKey()}; 362 } 363 static unsigned getHashValue(const VTableSlotSummary &I) { 364 return DenseMapInfo<StringRef>::getHashValue(I.TypeID) ^ 365 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 366 } 367 static bool isEqual(const VTableSlotSummary &LHS, 368 const VTableSlotSummary &RHS) { 369 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 370 } 371 }; 372 373 } // end namespace llvm 374 375 namespace { 376 377 // Returns true if the function must be unreachable based on ValueInfo. 378 // 379 // In particular, identifies a function as unreachable in the following 380 // conditions 381 // 1) All summaries are live. 382 // 2) All function summaries indicate it's unreachable 383 bool mustBeUnreachableFunction(ValueInfo TheFnVI) { 384 if ((!TheFnVI) || TheFnVI.getSummaryList().empty()) { 385 // Returns false if ValueInfo is absent, or the summary list is empty 386 // (e.g., function declarations). 387 return false; 388 } 389 390 for (auto &Summary : TheFnVI.getSummaryList()) { 391 // Conservatively returns false if any non-live functions are seen. 392 // In general either all summaries should be live or all should be dead. 393 if (!Summary->isLive()) 394 return false; 395 if (auto *FS = dyn_cast<FunctionSummary>(Summary.get())) { 396 if (!FS->fflags().MustBeUnreachable) 397 return false; 398 } 399 // Do nothing if a non-function has the same GUID (which is rare). 400 // This is correct since non-function summaries are not relevant. 401 } 402 // All function summaries are live and all of them agree that the function is 403 // unreachble. 404 return true; 405 } 406 407 // A virtual call site. VTable is the loaded virtual table pointer, and CS is 408 // the indirect virtual call. 409 struct VirtualCallSite { 410 Value *VTable = nullptr; 411 CallBase &CB; 412 413 // If non-null, this field points to the associated unsafe use count stored in 414 // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description 415 // of that field for details. 416 unsigned *NumUnsafeUses = nullptr; 417 418 void 419 emitRemark(const StringRef OptName, const StringRef TargetName, 420 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) { 421 Function *F = CB.getCaller(); 422 DebugLoc DLoc = CB.getDebugLoc(); 423 BasicBlock *Block = CB.getParent(); 424 425 using namespace ore; 426 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, OptName, DLoc, Block) 427 << NV("Optimization", OptName) 428 << ": devirtualized a call to " 429 << NV("FunctionName", TargetName)); 430 } 431 432 void replaceAndErase( 433 const StringRef OptName, const StringRef TargetName, bool RemarksEnabled, 434 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 435 Value *New) { 436 if (RemarksEnabled) 437 emitRemark(OptName, TargetName, OREGetter); 438 CB.replaceAllUsesWith(New); 439 if (auto *II = dyn_cast<InvokeInst>(&CB)) { 440 BranchInst::Create(II->getNormalDest(), &CB); 441 II->getUnwindDest()->removePredecessor(II->getParent()); 442 } 443 CB.eraseFromParent(); 444 // This use is no longer unsafe. 445 if (NumUnsafeUses) 446 --*NumUnsafeUses; 447 } 448 }; 449 450 // Call site information collected for a specific VTableSlot and possibly a list 451 // of constant integer arguments. The grouping by arguments is handled by the 452 // VTableSlotInfo class. 453 struct CallSiteInfo { 454 /// The set of call sites for this slot. Used during regular LTO and the 455 /// import phase of ThinLTO (as well as the export phase of ThinLTO for any 456 /// call sites that appear in the merged module itself); in each of these 457 /// cases we are directly operating on the call sites at the IR level. 458 std::vector<VirtualCallSite> CallSites; 459 460 /// Whether all call sites represented by this CallSiteInfo, including those 461 /// in summaries, have been devirtualized. This starts off as true because a 462 /// default constructed CallSiteInfo represents no call sites. 463 bool AllCallSitesDevirted = true; 464 465 // These fields are used during the export phase of ThinLTO and reflect 466 // information collected from function summaries. 467 468 /// Whether any function summary contains an llvm.assume(llvm.type.test) for 469 /// this slot. 470 bool SummaryHasTypeTestAssumeUsers = false; 471 472 /// CFI-specific: a vector containing the list of function summaries that use 473 /// the llvm.type.checked.load intrinsic and therefore will require 474 /// resolutions for llvm.type.test in order to implement CFI checks if 475 /// devirtualization was unsuccessful. If devirtualization was successful, the 476 /// pass will clear this vector by calling markDevirt(). If at the end of the 477 /// pass the vector is non-empty, we will need to add a use of llvm.type.test 478 /// to each of the function summaries in the vector. 479 std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers; 480 std::vector<FunctionSummary *> SummaryTypeTestAssumeUsers; 481 482 bool isExported() const { 483 return SummaryHasTypeTestAssumeUsers || 484 !SummaryTypeCheckedLoadUsers.empty(); 485 } 486 487 void addSummaryTypeCheckedLoadUser(FunctionSummary *FS) { 488 SummaryTypeCheckedLoadUsers.push_back(FS); 489 AllCallSitesDevirted = false; 490 } 491 492 void addSummaryTypeTestAssumeUser(FunctionSummary *FS) { 493 SummaryTypeTestAssumeUsers.push_back(FS); 494 SummaryHasTypeTestAssumeUsers = true; 495 AllCallSitesDevirted = false; 496 } 497 498 void markDevirt() { 499 AllCallSitesDevirted = true; 500 501 // As explained in the comment for SummaryTypeCheckedLoadUsers. 502 SummaryTypeCheckedLoadUsers.clear(); 503 } 504 }; 505 506 // Call site information collected for a specific VTableSlot. 507 struct VTableSlotInfo { 508 // The set of call sites which do not have all constant integer arguments 509 // (excluding "this"). 510 CallSiteInfo CSInfo; 511 512 // The set of call sites with all constant integer arguments (excluding 513 // "this"), grouped by argument list. 514 std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo; 515 516 void addCallSite(Value *VTable, CallBase &CB, unsigned *NumUnsafeUses); 517 518 private: 519 CallSiteInfo &findCallSiteInfo(CallBase &CB); 520 }; 521 522 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallBase &CB) { 523 std::vector<uint64_t> Args; 524 auto *CBType = dyn_cast<IntegerType>(CB.getType()); 525 if (!CBType || CBType->getBitWidth() > 64 || CB.arg_empty()) 526 return CSInfo; 527 for (auto &&Arg : drop_begin(CB.args())) { 528 auto *CI = dyn_cast<ConstantInt>(Arg); 529 if (!CI || CI->getBitWidth() > 64) 530 return CSInfo; 531 Args.push_back(CI->getZExtValue()); 532 } 533 return ConstCSInfo[Args]; 534 } 535 536 void VTableSlotInfo::addCallSite(Value *VTable, CallBase &CB, 537 unsigned *NumUnsafeUses) { 538 auto &CSI = findCallSiteInfo(CB); 539 CSI.AllCallSitesDevirted = false; 540 CSI.CallSites.push_back({VTable, CB, NumUnsafeUses}); 541 } 542 543 struct DevirtModule { 544 Module &M; 545 function_ref<AAResults &(Function &)> AARGetter; 546 function_ref<DominatorTree &(Function &)> LookupDomTree; 547 548 ModuleSummaryIndex *ExportSummary; 549 const ModuleSummaryIndex *ImportSummary; 550 551 IntegerType *Int8Ty; 552 PointerType *Int8PtrTy; 553 IntegerType *Int32Ty; 554 IntegerType *Int64Ty; 555 IntegerType *IntPtrTy; 556 /// Sizeless array type, used for imported vtables. This provides a signal 557 /// to analyzers that these imports may alias, as they do for example 558 /// when multiple unique return values occur in the same vtable. 559 ArrayType *Int8Arr0Ty; 560 561 bool RemarksEnabled; 562 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter; 563 564 MapVector<VTableSlot, VTableSlotInfo> CallSlots; 565 566 // Calls that have already been optimized. We may add a call to multiple 567 // VTableSlotInfos if vtable loads are coalesced and need to make sure not to 568 // optimize a call more than once. 569 SmallPtrSet<CallBase *, 8> OptimizedCalls; 570 571 // This map keeps track of the number of "unsafe" uses of a loaded function 572 // pointer. The key is the associated llvm.type.test intrinsic call generated 573 // by this pass. An unsafe use is one that calls the loaded function pointer 574 // directly. Every time we eliminate an unsafe use (for example, by 575 // devirtualizing it or by applying virtual constant propagation), we 576 // decrement the value stored in this map. If a value reaches zero, we can 577 // eliminate the type check by RAUWing the associated llvm.type.test call with 578 // true. 579 std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest; 580 PatternList FunctionsToSkip; 581 582 DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter, 583 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 584 function_ref<DominatorTree &(Function &)> LookupDomTree, 585 ModuleSummaryIndex *ExportSummary, 586 const ModuleSummaryIndex *ImportSummary) 587 : M(M), AARGetter(AARGetter), LookupDomTree(LookupDomTree), 588 ExportSummary(ExportSummary), ImportSummary(ImportSummary), 589 Int8Ty(Type::getInt8Ty(M.getContext())), 590 Int8PtrTy(Type::getInt8PtrTy(M.getContext())), 591 Int32Ty(Type::getInt32Ty(M.getContext())), 592 Int64Ty(Type::getInt64Ty(M.getContext())), 593 IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)), 594 Int8Arr0Ty(ArrayType::get(Type::getInt8Ty(M.getContext()), 0)), 595 RemarksEnabled(areRemarksEnabled()), OREGetter(OREGetter) { 596 assert(!(ExportSummary && ImportSummary)); 597 FunctionsToSkip.init(SkipFunctionNames); 598 } 599 600 bool areRemarksEnabled(); 601 602 void 603 scanTypeTestUsers(Function *TypeTestFunc, 604 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap); 605 void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc); 606 607 void buildTypeIdentifierMap( 608 std::vector<VTableBits> &Bits, 609 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap); 610 611 bool 612 tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot, 613 const std::set<TypeMemberInfo> &TypeMemberInfos, 614 uint64_t ByteOffset, 615 ModuleSummaryIndex *ExportSummary); 616 617 void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn, 618 bool &IsExported); 619 bool trySingleImplDevirt(ModuleSummaryIndex *ExportSummary, 620 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 621 VTableSlotInfo &SlotInfo, 622 WholeProgramDevirtResolution *Res); 623 624 void applyICallBranchFunnel(VTableSlotInfo &SlotInfo, Constant *JT, 625 bool &IsExported); 626 void tryICallBranchFunnel(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 627 VTableSlotInfo &SlotInfo, 628 WholeProgramDevirtResolution *Res, VTableSlot Slot); 629 630 bool tryEvaluateFunctionsWithArgs( 631 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 632 ArrayRef<uint64_t> Args); 633 634 void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 635 uint64_t TheRetVal); 636 bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 637 CallSiteInfo &CSInfo, 638 WholeProgramDevirtResolution::ByArg *Res); 639 640 // Returns the global symbol name that is used to export information about the 641 // given vtable slot and list of arguments. 642 std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args, 643 StringRef Name); 644 645 bool shouldExportConstantsAsAbsoluteSymbols(); 646 647 // This function is called during the export phase to create a symbol 648 // definition containing information about the given vtable slot and list of 649 // arguments. 650 void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 651 Constant *C); 652 void exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 653 uint32_t Const, uint32_t &Storage); 654 655 // This function is called during the import phase to create a reference to 656 // the symbol definition created during the export phase. 657 Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 658 StringRef Name); 659 Constant *importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 660 StringRef Name, IntegerType *IntTy, 661 uint32_t Storage); 662 663 Constant *getMemberAddr(const TypeMemberInfo *M); 664 665 void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne, 666 Constant *UniqueMemberAddr); 667 bool tryUniqueRetValOpt(unsigned BitWidth, 668 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 669 CallSiteInfo &CSInfo, 670 WholeProgramDevirtResolution::ByArg *Res, 671 VTableSlot Slot, ArrayRef<uint64_t> Args); 672 673 void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 674 Constant *Byte, Constant *Bit); 675 bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 676 VTableSlotInfo &SlotInfo, 677 WholeProgramDevirtResolution *Res, VTableSlot Slot); 678 679 void rebuildGlobal(VTableBits &B); 680 681 // Apply the summary resolution for Slot to all virtual calls in SlotInfo. 682 void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo); 683 684 // If we were able to eliminate all unsafe uses for a type checked load, 685 // eliminate the associated type tests by replacing them with true. 686 void removeRedundantTypeTests(); 687 688 bool run(); 689 690 // Look up the corresponding ValueInfo entry of `TheFn` in `ExportSummary`. 691 // 692 // Caller guarantees that `ExportSummary` is not nullptr. 693 static ValueInfo lookUpFunctionValueInfo(Function *TheFn, 694 ModuleSummaryIndex *ExportSummary); 695 696 // Returns true if the function definition must be unreachable. 697 // 698 // Note if this helper function returns true, `F` is guaranteed 699 // to be unreachable; if it returns false, `F` might still 700 // be unreachable but not covered by this helper function. 701 // 702 // Implementation-wise, if function definition is present, IR is analyzed; if 703 // not, look up function flags from ExportSummary as a fallback. 704 static bool mustBeUnreachableFunction(Function *const F, 705 ModuleSummaryIndex *ExportSummary); 706 707 // Lower the module using the action and summary passed as command line 708 // arguments. For testing purposes only. 709 static bool 710 runForTesting(Module &M, function_ref<AAResults &(Function &)> AARGetter, 711 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 712 function_ref<DominatorTree &(Function &)> LookupDomTree); 713 }; 714 715 struct DevirtIndex { 716 ModuleSummaryIndex &ExportSummary; 717 // The set in which to record GUIDs exported from their module by 718 // devirtualization, used by client to ensure they are not internalized. 719 std::set<GlobalValue::GUID> &ExportedGUIDs; 720 // A map in which to record the information necessary to locate the WPD 721 // resolution for local targets in case they are exported by cross module 722 // importing. 723 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap; 724 725 MapVector<VTableSlotSummary, VTableSlotInfo> CallSlots; 726 727 PatternList FunctionsToSkip; 728 729 DevirtIndex( 730 ModuleSummaryIndex &ExportSummary, 731 std::set<GlobalValue::GUID> &ExportedGUIDs, 732 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) 733 : ExportSummary(ExportSummary), ExportedGUIDs(ExportedGUIDs), 734 LocalWPDTargetsMap(LocalWPDTargetsMap) { 735 FunctionsToSkip.init(SkipFunctionNames); 736 } 737 738 bool tryFindVirtualCallTargets(std::vector<ValueInfo> &TargetsForSlot, 739 const TypeIdCompatibleVtableInfo TIdInfo, 740 uint64_t ByteOffset); 741 742 bool trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 743 VTableSlotSummary &SlotSummary, 744 VTableSlotInfo &SlotInfo, 745 WholeProgramDevirtResolution *Res, 746 std::set<ValueInfo> &DevirtTargets); 747 748 void run(); 749 }; 750 } // end anonymous namespace 751 752 PreservedAnalyses WholeProgramDevirtPass::run(Module &M, 753 ModuleAnalysisManager &AM) { 754 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 755 auto AARGetter = [&](Function &F) -> AAResults & { 756 return FAM.getResult<AAManager>(F); 757 }; 758 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { 759 return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F); 760 }; 761 auto LookupDomTree = [&FAM](Function &F) -> DominatorTree & { 762 return FAM.getResult<DominatorTreeAnalysis>(F); 763 }; 764 if (UseCommandLine) { 765 if (DevirtModule::runForTesting(M, AARGetter, OREGetter, LookupDomTree)) 766 return PreservedAnalyses::all(); 767 return PreservedAnalyses::none(); 768 } 769 if (!DevirtModule(M, AARGetter, OREGetter, LookupDomTree, ExportSummary, 770 ImportSummary) 771 .run()) 772 return PreservedAnalyses::all(); 773 return PreservedAnalyses::none(); 774 } 775 776 namespace llvm { 777 // Enable whole program visibility if enabled by client (e.g. linker) or 778 // internal option, and not force disabled. 779 bool hasWholeProgramVisibility(bool WholeProgramVisibilityEnabledInLTO) { 780 return (WholeProgramVisibilityEnabledInLTO || WholeProgramVisibility) && 781 !DisableWholeProgramVisibility; 782 } 783 784 /// If whole program visibility asserted, then upgrade all public vcall 785 /// visibility metadata on vtable definitions to linkage unit visibility in 786 /// Module IR (for regular or hybrid LTO). 787 void updateVCallVisibilityInModule( 788 Module &M, bool WholeProgramVisibilityEnabledInLTO, 789 const DenseSet<GlobalValue::GUID> &DynamicExportSymbols) { 790 if (!hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO)) 791 return; 792 for (GlobalVariable &GV : M.globals()) { 793 // Add linkage unit visibility to any variable with type metadata, which are 794 // the vtable definitions. We won't have an existing vcall_visibility 795 // metadata on vtable definitions with public visibility. 796 if (GV.hasMetadata(LLVMContext::MD_type) && 797 GV.getVCallVisibility() == GlobalObject::VCallVisibilityPublic && 798 // Don't upgrade the visibility for symbols exported to the dynamic 799 // linker, as we have no information on their eventual use. 800 !DynamicExportSymbols.count(GV.getGUID())) 801 GV.setVCallVisibilityMetadata(GlobalObject::VCallVisibilityLinkageUnit); 802 } 803 } 804 805 void updatePublicTypeTestCalls(Module &M, 806 bool WholeProgramVisibilityEnabledInLTO) { 807 Function *PublicTypeTestFunc = 808 M.getFunction(Intrinsic::getName(Intrinsic::public_type_test)); 809 if (!PublicTypeTestFunc) 810 return; 811 if (hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO)) { 812 Function *TypeTestFunc = 813 Intrinsic::getDeclaration(&M, Intrinsic::type_test); 814 for (Use &U : make_early_inc_range(PublicTypeTestFunc->uses())) { 815 auto *CI = cast<CallInst>(U.getUser()); 816 auto *NewCI = CallInst::Create( 817 TypeTestFunc, {CI->getArgOperand(0), CI->getArgOperand(1)}, None, "", 818 CI); 819 CI->replaceAllUsesWith(NewCI); 820 CI->eraseFromParent(); 821 } 822 } else { 823 auto *True = ConstantInt::getTrue(M.getContext()); 824 for (Use &U : make_early_inc_range(PublicTypeTestFunc->uses())) { 825 auto *CI = cast<CallInst>(U.getUser()); 826 CI->replaceAllUsesWith(True); 827 CI->eraseFromParent(); 828 } 829 } 830 } 831 832 /// If whole program visibility asserted, then upgrade all public vcall 833 /// visibility metadata on vtable definition summaries to linkage unit 834 /// visibility in Module summary index (for ThinLTO). 835 void updateVCallVisibilityInIndex( 836 ModuleSummaryIndex &Index, bool WholeProgramVisibilityEnabledInLTO, 837 const DenseSet<GlobalValue::GUID> &DynamicExportSymbols) { 838 if (!hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO)) 839 return; 840 for (auto &P : Index) { 841 // Don't upgrade the visibility for symbols exported to the dynamic 842 // linker, as we have no information on their eventual use. 843 if (DynamicExportSymbols.count(P.first)) 844 continue; 845 for (auto &S : P.second.SummaryList) { 846 auto *GVar = dyn_cast<GlobalVarSummary>(S.get()); 847 if (!GVar || 848 GVar->getVCallVisibility() != GlobalObject::VCallVisibilityPublic) 849 continue; 850 GVar->setVCallVisibility(GlobalObject::VCallVisibilityLinkageUnit); 851 } 852 } 853 } 854 855 void runWholeProgramDevirtOnIndex( 856 ModuleSummaryIndex &Summary, std::set<GlobalValue::GUID> &ExportedGUIDs, 857 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 858 DevirtIndex(Summary, ExportedGUIDs, LocalWPDTargetsMap).run(); 859 } 860 861 void updateIndexWPDForExports( 862 ModuleSummaryIndex &Summary, 863 function_ref<bool(StringRef, ValueInfo)> isExported, 864 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 865 for (auto &T : LocalWPDTargetsMap) { 866 auto &VI = T.first; 867 // This was enforced earlier during trySingleImplDevirt. 868 assert(VI.getSummaryList().size() == 1 && 869 "Devirt of local target has more than one copy"); 870 auto &S = VI.getSummaryList()[0]; 871 if (!isExported(S->modulePath(), VI)) 872 continue; 873 874 // It's been exported by a cross module import. 875 for (auto &SlotSummary : T.second) { 876 auto *TIdSum = Summary.getTypeIdSummary(SlotSummary.TypeID); 877 assert(TIdSum); 878 auto WPDRes = TIdSum->WPDRes.find(SlotSummary.ByteOffset); 879 assert(WPDRes != TIdSum->WPDRes.end()); 880 WPDRes->second.SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 881 WPDRes->second.SingleImplName, 882 Summary.getModuleHash(S->modulePath())); 883 } 884 } 885 } 886 887 } // end namespace llvm 888 889 static Error checkCombinedSummaryForTesting(ModuleSummaryIndex *Summary) { 890 // Check that summary index contains regular LTO module when performing 891 // export to prevent occasional use of index from pure ThinLTO compilation 892 // (-fno-split-lto-module). This kind of summary index is passed to 893 // DevirtIndex::run, not to DevirtModule::run used by opt/runForTesting. 894 const auto &ModPaths = Summary->modulePaths(); 895 if (ClSummaryAction != PassSummaryAction::Import && 896 ModPaths.find(ModuleSummaryIndex::getRegularLTOModuleName()) == 897 ModPaths.end()) 898 return createStringError( 899 errc::invalid_argument, 900 "combined summary should contain Regular LTO module"); 901 return ErrorSuccess(); 902 } 903 904 bool DevirtModule::runForTesting( 905 Module &M, function_ref<AAResults &(Function &)> AARGetter, 906 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 907 function_ref<DominatorTree &(Function &)> LookupDomTree) { 908 std::unique_ptr<ModuleSummaryIndex> Summary = 909 std::make_unique<ModuleSummaryIndex>(/*HaveGVs=*/false); 910 911 // Handle the command-line summary arguments. This code is for testing 912 // purposes only, so we handle errors directly. 913 if (!ClReadSummary.empty()) { 914 ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary + 915 ": "); 916 auto ReadSummaryFile = 917 ExitOnErr(errorOrToExpected(MemoryBuffer::getFile(ClReadSummary))); 918 if (Expected<std::unique_ptr<ModuleSummaryIndex>> SummaryOrErr = 919 getModuleSummaryIndex(*ReadSummaryFile)) { 920 Summary = std::move(*SummaryOrErr); 921 ExitOnErr(checkCombinedSummaryForTesting(Summary.get())); 922 } else { 923 // Try YAML if we've failed with bitcode. 924 consumeError(SummaryOrErr.takeError()); 925 yaml::Input In(ReadSummaryFile->getBuffer()); 926 In >> *Summary; 927 ExitOnErr(errorCodeToError(In.error())); 928 } 929 } 930 931 bool Changed = 932 DevirtModule(M, AARGetter, OREGetter, LookupDomTree, 933 ClSummaryAction == PassSummaryAction::Export ? Summary.get() 934 : nullptr, 935 ClSummaryAction == PassSummaryAction::Import ? Summary.get() 936 : nullptr) 937 .run(); 938 939 if (!ClWriteSummary.empty()) { 940 ExitOnError ExitOnErr( 941 "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": "); 942 std::error_code EC; 943 if (StringRef(ClWriteSummary).endswith(".bc")) { 944 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_None); 945 ExitOnErr(errorCodeToError(EC)); 946 writeIndexToFile(*Summary, OS); 947 } else { 948 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_TextWithCRLF); 949 ExitOnErr(errorCodeToError(EC)); 950 yaml::Output Out(OS); 951 Out << *Summary; 952 } 953 } 954 955 return Changed; 956 } 957 958 void DevirtModule::buildTypeIdentifierMap( 959 std::vector<VTableBits> &Bits, 960 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 961 DenseMap<GlobalVariable *, VTableBits *> GVToBits; 962 Bits.reserve(M.getGlobalList().size()); 963 SmallVector<MDNode *, 2> Types; 964 for (GlobalVariable &GV : M.globals()) { 965 Types.clear(); 966 GV.getMetadata(LLVMContext::MD_type, Types); 967 if (GV.isDeclaration() || Types.empty()) 968 continue; 969 970 VTableBits *&BitsPtr = GVToBits[&GV]; 971 if (!BitsPtr) { 972 Bits.emplace_back(); 973 Bits.back().GV = &GV; 974 Bits.back().ObjectSize = 975 M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType()); 976 BitsPtr = &Bits.back(); 977 } 978 979 for (MDNode *Type : Types) { 980 auto TypeID = Type->getOperand(1).get(); 981 982 uint64_t Offset = 983 cast<ConstantInt>( 984 cast<ConstantAsMetadata>(Type->getOperand(0))->getValue()) 985 ->getZExtValue(); 986 987 TypeIdMap[TypeID].insert({BitsPtr, Offset}); 988 } 989 } 990 } 991 992 bool DevirtModule::tryFindVirtualCallTargets( 993 std::vector<VirtualCallTarget> &TargetsForSlot, 994 const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset, 995 ModuleSummaryIndex *ExportSummary) { 996 for (const TypeMemberInfo &TM : TypeMemberInfos) { 997 if (!TM.Bits->GV->isConstant()) 998 return false; 999 1000 // We cannot perform whole program devirtualization analysis on a vtable 1001 // with public LTO visibility. 1002 if (TM.Bits->GV->getVCallVisibility() == 1003 GlobalObject::VCallVisibilityPublic) 1004 return false; 1005 1006 Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(), 1007 TM.Offset + ByteOffset, M); 1008 if (!Ptr) 1009 return false; 1010 1011 auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts()); 1012 if (!Fn) 1013 return false; 1014 1015 if (FunctionsToSkip.match(Fn->getName())) 1016 return false; 1017 1018 // We can disregard __cxa_pure_virtual as a possible call target, as 1019 // calls to pure virtuals are UB. 1020 if (Fn->getName() == "__cxa_pure_virtual") 1021 continue; 1022 1023 // We can disregard unreachable functions as possible call targets, as 1024 // unreachable functions shouldn't be called. 1025 if (mustBeUnreachableFunction(Fn, ExportSummary)) 1026 continue; 1027 1028 TargetsForSlot.push_back({Fn, &TM}); 1029 } 1030 1031 // Give up if we couldn't find any targets. 1032 return !TargetsForSlot.empty(); 1033 } 1034 1035 bool DevirtIndex::tryFindVirtualCallTargets( 1036 std::vector<ValueInfo> &TargetsForSlot, const TypeIdCompatibleVtableInfo TIdInfo, 1037 uint64_t ByteOffset) { 1038 for (const TypeIdOffsetVtableInfo &P : TIdInfo) { 1039 // Find a representative copy of the vtable initializer. 1040 // We can have multiple available_externally, linkonce_odr and weak_odr 1041 // vtable initializers. We can also have multiple external vtable 1042 // initializers in the case of comdats, which we cannot check here. 1043 // The linker should give an error in this case. 1044 // 1045 // Also, handle the case of same-named local Vtables with the same path 1046 // and therefore the same GUID. This can happen if there isn't enough 1047 // distinguishing path when compiling the source file. In that case we 1048 // conservatively return false early. 1049 const GlobalVarSummary *VS = nullptr; 1050 bool LocalFound = false; 1051 for (auto &S : P.VTableVI.getSummaryList()) { 1052 if (GlobalValue::isLocalLinkage(S->linkage())) { 1053 if (LocalFound) 1054 return false; 1055 LocalFound = true; 1056 } 1057 auto *CurVS = cast<GlobalVarSummary>(S->getBaseObject()); 1058 if (!CurVS->vTableFuncs().empty() || 1059 // Previously clang did not attach the necessary type metadata to 1060 // available_externally vtables, in which case there would not 1061 // be any vtable functions listed in the summary and we need 1062 // to treat this case conservatively (in case the bitcode is old). 1063 // However, we will also not have any vtable functions in the 1064 // case of a pure virtual base class. In that case we do want 1065 // to set VS to avoid treating it conservatively. 1066 !GlobalValue::isAvailableExternallyLinkage(S->linkage())) { 1067 VS = CurVS; 1068 // We cannot perform whole program devirtualization analysis on a vtable 1069 // with public LTO visibility. 1070 if (VS->getVCallVisibility() == GlobalObject::VCallVisibilityPublic) 1071 return false; 1072 } 1073 } 1074 // There will be no VS if all copies are available_externally having no 1075 // type metadata. In that case we can't safely perform WPD. 1076 if (!VS) 1077 return false; 1078 if (!VS->isLive()) 1079 continue; 1080 for (auto VTP : VS->vTableFuncs()) { 1081 if (VTP.VTableOffset != P.AddressPointOffset + ByteOffset) 1082 continue; 1083 1084 if (mustBeUnreachableFunction(VTP.FuncVI)) 1085 continue; 1086 1087 TargetsForSlot.push_back(VTP.FuncVI); 1088 } 1089 } 1090 1091 // Give up if we couldn't find any targets. 1092 return !TargetsForSlot.empty(); 1093 } 1094 1095 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo, 1096 Constant *TheFn, bool &IsExported) { 1097 // Don't devirtualize function if we're told to skip it 1098 // in -wholeprogramdevirt-skip. 1099 if (FunctionsToSkip.match(TheFn->stripPointerCasts()->getName())) 1100 return; 1101 auto Apply = [&](CallSiteInfo &CSInfo) { 1102 for (auto &&VCallSite : CSInfo.CallSites) { 1103 if (!OptimizedCalls.insert(&VCallSite.CB).second) 1104 continue; 1105 1106 if (RemarksEnabled) 1107 VCallSite.emitRemark("single-impl", 1108 TheFn->stripPointerCasts()->getName(), OREGetter); 1109 NumSingleImpl++; 1110 auto &CB = VCallSite.CB; 1111 assert(!CB.getCalledFunction() && "devirtualizing direct call?"); 1112 IRBuilder<> Builder(&CB); 1113 Value *Callee = 1114 Builder.CreateBitCast(TheFn, CB.getCalledOperand()->getType()); 1115 1116 // If trap checking is enabled, add support to compare the virtual 1117 // function pointer to the devirtualized target. In case of a mismatch, 1118 // perform a debug trap. 1119 if (DevirtCheckMode == WPDCheckMode::Trap) { 1120 auto *Cond = Builder.CreateICmpNE(CB.getCalledOperand(), Callee); 1121 Instruction *ThenTerm = 1122 SplitBlockAndInsertIfThen(Cond, &CB, /*Unreachable=*/false); 1123 Builder.SetInsertPoint(ThenTerm); 1124 Function *TrapFn = Intrinsic::getDeclaration(&M, Intrinsic::debugtrap); 1125 auto *CallTrap = Builder.CreateCall(TrapFn); 1126 CallTrap->setDebugLoc(CB.getDebugLoc()); 1127 } 1128 1129 // If fallback checking is enabled, add support to compare the virtual 1130 // function pointer to the devirtualized target. In case of a mismatch, 1131 // fall back to indirect call. 1132 if (DevirtCheckMode == WPDCheckMode::Fallback) { 1133 MDNode *Weights = 1134 MDBuilder(M.getContext()).createBranchWeights((1U << 20) - 1, 1); 1135 // Version the indirect call site. If the called value is equal to the 1136 // given callee, 'NewInst' will be executed, otherwise the original call 1137 // site will be executed. 1138 CallBase &NewInst = versionCallSite(CB, Callee, Weights); 1139 NewInst.setCalledOperand(Callee); 1140 // Since the new call site is direct, we must clear metadata that 1141 // is only appropriate for indirect calls. This includes !prof and 1142 // !callees metadata. 1143 NewInst.setMetadata(LLVMContext::MD_prof, nullptr); 1144 NewInst.setMetadata(LLVMContext::MD_callees, nullptr); 1145 // Additionally, we should remove them from the fallback indirect call, 1146 // so that we don't attempt to perform indirect call promotion later. 1147 CB.setMetadata(LLVMContext::MD_prof, nullptr); 1148 CB.setMetadata(LLVMContext::MD_callees, nullptr); 1149 } 1150 1151 // In either trapping or non-checking mode, devirtualize original call. 1152 else { 1153 // Devirtualize unconditionally. 1154 CB.setCalledOperand(Callee); 1155 // Since the call site is now direct, we must clear metadata that 1156 // is only appropriate for indirect calls. This includes !prof and 1157 // !callees metadata. 1158 CB.setMetadata(LLVMContext::MD_prof, nullptr); 1159 CB.setMetadata(LLVMContext::MD_callees, nullptr); 1160 } 1161 1162 // This use is no longer unsafe. 1163 if (VCallSite.NumUnsafeUses) 1164 --*VCallSite.NumUnsafeUses; 1165 } 1166 if (CSInfo.isExported()) 1167 IsExported = true; 1168 CSInfo.markDevirt(); 1169 }; 1170 Apply(SlotInfo.CSInfo); 1171 for (auto &P : SlotInfo.ConstCSInfo) 1172 Apply(P.second); 1173 } 1174 1175 static bool AddCalls(VTableSlotInfo &SlotInfo, const ValueInfo &Callee) { 1176 // We can't add calls if we haven't seen a definition 1177 if (Callee.getSummaryList().empty()) 1178 return false; 1179 1180 // Insert calls into the summary index so that the devirtualized targets 1181 // are eligible for import. 1182 // FIXME: Annotate type tests with hotness. For now, mark these as hot 1183 // to better ensure we have the opportunity to inline them. 1184 bool IsExported = false; 1185 auto &S = Callee.getSummaryList()[0]; 1186 CalleeInfo CI(CalleeInfo::HotnessType::Hot, /* RelBF = */ 0); 1187 auto AddCalls = [&](CallSiteInfo &CSInfo) { 1188 for (auto *FS : CSInfo.SummaryTypeCheckedLoadUsers) { 1189 FS->addCall({Callee, CI}); 1190 IsExported |= S->modulePath() != FS->modulePath(); 1191 } 1192 for (auto *FS : CSInfo.SummaryTypeTestAssumeUsers) { 1193 FS->addCall({Callee, CI}); 1194 IsExported |= S->modulePath() != FS->modulePath(); 1195 } 1196 }; 1197 AddCalls(SlotInfo.CSInfo); 1198 for (auto &P : SlotInfo.ConstCSInfo) 1199 AddCalls(P.second); 1200 return IsExported; 1201 } 1202 1203 bool DevirtModule::trySingleImplDevirt( 1204 ModuleSummaryIndex *ExportSummary, 1205 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1206 WholeProgramDevirtResolution *Res) { 1207 // See if the program contains a single implementation of this virtual 1208 // function. 1209 Function *TheFn = TargetsForSlot[0].Fn; 1210 for (auto &&Target : TargetsForSlot) 1211 if (TheFn != Target.Fn) 1212 return false; 1213 1214 // If so, update each call site to call that implementation directly. 1215 if (RemarksEnabled || AreStatisticsEnabled()) 1216 TargetsForSlot[0].WasDevirt = true; 1217 1218 bool IsExported = false; 1219 applySingleImplDevirt(SlotInfo, TheFn, IsExported); 1220 if (!IsExported) 1221 return false; 1222 1223 // If the only implementation has local linkage, we must promote to external 1224 // to make it visible to thin LTO objects. We can only get here during the 1225 // ThinLTO export phase. 1226 if (TheFn->hasLocalLinkage()) { 1227 std::string NewName = (TheFn->getName() + ".llvm.merged").str(); 1228 1229 // Since we are renaming the function, any comdats with the same name must 1230 // also be renamed. This is required when targeting COFF, as the comdat name 1231 // must match one of the names of the symbols in the comdat. 1232 if (Comdat *C = TheFn->getComdat()) { 1233 if (C->getName() == TheFn->getName()) { 1234 Comdat *NewC = M.getOrInsertComdat(NewName); 1235 NewC->setSelectionKind(C->getSelectionKind()); 1236 for (GlobalObject &GO : M.global_objects()) 1237 if (GO.getComdat() == C) 1238 GO.setComdat(NewC); 1239 } 1240 } 1241 1242 TheFn->setLinkage(GlobalValue::ExternalLinkage); 1243 TheFn->setVisibility(GlobalValue::HiddenVisibility); 1244 TheFn->setName(NewName); 1245 } 1246 if (ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFn->getGUID())) 1247 // Any needed promotion of 'TheFn' has already been done during 1248 // LTO unit split, so we can ignore return value of AddCalls. 1249 AddCalls(SlotInfo, TheFnVI); 1250 1251 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1252 Res->SingleImplName = std::string(TheFn->getName()); 1253 1254 return true; 1255 } 1256 1257 bool DevirtIndex::trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 1258 VTableSlotSummary &SlotSummary, 1259 VTableSlotInfo &SlotInfo, 1260 WholeProgramDevirtResolution *Res, 1261 std::set<ValueInfo> &DevirtTargets) { 1262 // See if the program contains a single implementation of this virtual 1263 // function. 1264 auto TheFn = TargetsForSlot[0]; 1265 for (auto &&Target : TargetsForSlot) 1266 if (TheFn != Target) 1267 return false; 1268 1269 // Don't devirtualize if we don't have target definition. 1270 auto Size = TheFn.getSummaryList().size(); 1271 if (!Size) 1272 return false; 1273 1274 // Don't devirtualize function if we're told to skip it 1275 // in -wholeprogramdevirt-skip. 1276 if (FunctionsToSkip.match(TheFn.name())) 1277 return false; 1278 1279 // If the summary list contains multiple summaries where at least one is 1280 // a local, give up, as we won't know which (possibly promoted) name to use. 1281 for (auto &S : TheFn.getSummaryList()) 1282 if (GlobalValue::isLocalLinkage(S->linkage()) && Size > 1) 1283 return false; 1284 1285 // Collect functions devirtualized at least for one call site for stats. 1286 if (PrintSummaryDevirt || AreStatisticsEnabled()) 1287 DevirtTargets.insert(TheFn); 1288 1289 auto &S = TheFn.getSummaryList()[0]; 1290 bool IsExported = AddCalls(SlotInfo, TheFn); 1291 if (IsExported) 1292 ExportedGUIDs.insert(TheFn.getGUID()); 1293 1294 // Record in summary for use in devirtualization during the ThinLTO import 1295 // step. 1296 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1297 if (GlobalValue::isLocalLinkage(S->linkage())) { 1298 if (IsExported) 1299 // If target is a local function and we are exporting it by 1300 // devirtualizing a call in another module, we need to record the 1301 // promoted name. 1302 Res->SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 1303 TheFn.name(), ExportSummary.getModuleHash(S->modulePath())); 1304 else { 1305 LocalWPDTargetsMap[TheFn].push_back(SlotSummary); 1306 Res->SingleImplName = std::string(TheFn.name()); 1307 } 1308 } else 1309 Res->SingleImplName = std::string(TheFn.name()); 1310 1311 // Name will be empty if this thin link driven off of serialized combined 1312 // index (e.g. llvm-lto). However, WPD is not supported/invoked for the 1313 // legacy LTO API anyway. 1314 assert(!Res->SingleImplName.empty()); 1315 1316 return true; 1317 } 1318 1319 void DevirtModule::tryICallBranchFunnel( 1320 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1321 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1322 Triple T(M.getTargetTriple()); 1323 if (T.getArch() != Triple::x86_64) 1324 return; 1325 1326 if (TargetsForSlot.size() > ClThreshold) 1327 return; 1328 1329 bool HasNonDevirt = !SlotInfo.CSInfo.AllCallSitesDevirted; 1330 if (!HasNonDevirt) 1331 for (auto &P : SlotInfo.ConstCSInfo) 1332 if (!P.second.AllCallSitesDevirted) { 1333 HasNonDevirt = true; 1334 break; 1335 } 1336 1337 if (!HasNonDevirt) 1338 return; 1339 1340 FunctionType *FT = 1341 FunctionType::get(Type::getVoidTy(M.getContext()), {Int8PtrTy}, true); 1342 Function *JT; 1343 if (isa<MDString>(Slot.TypeID)) { 1344 JT = Function::Create(FT, Function::ExternalLinkage, 1345 M.getDataLayout().getProgramAddressSpace(), 1346 getGlobalName(Slot, {}, "branch_funnel"), &M); 1347 JT->setVisibility(GlobalValue::HiddenVisibility); 1348 } else { 1349 JT = Function::Create(FT, Function::InternalLinkage, 1350 M.getDataLayout().getProgramAddressSpace(), 1351 "branch_funnel", &M); 1352 } 1353 JT->addParamAttr(0, Attribute::Nest); 1354 1355 std::vector<Value *> JTArgs; 1356 JTArgs.push_back(JT->arg_begin()); 1357 for (auto &T : TargetsForSlot) { 1358 JTArgs.push_back(getMemberAddr(T.TM)); 1359 JTArgs.push_back(T.Fn); 1360 } 1361 1362 BasicBlock *BB = BasicBlock::Create(M.getContext(), "", JT, nullptr); 1363 Function *Intr = 1364 Intrinsic::getDeclaration(&M, llvm::Intrinsic::icall_branch_funnel, {}); 1365 1366 auto *CI = CallInst::Create(Intr, JTArgs, "", BB); 1367 CI->setTailCallKind(CallInst::TCK_MustTail); 1368 ReturnInst::Create(M.getContext(), nullptr, BB); 1369 1370 bool IsExported = false; 1371 applyICallBranchFunnel(SlotInfo, JT, IsExported); 1372 if (IsExported) 1373 Res->TheKind = WholeProgramDevirtResolution::BranchFunnel; 1374 } 1375 1376 void DevirtModule::applyICallBranchFunnel(VTableSlotInfo &SlotInfo, 1377 Constant *JT, bool &IsExported) { 1378 auto Apply = [&](CallSiteInfo &CSInfo) { 1379 if (CSInfo.isExported()) 1380 IsExported = true; 1381 if (CSInfo.AllCallSitesDevirted) 1382 return; 1383 for (auto &&VCallSite : CSInfo.CallSites) { 1384 CallBase &CB = VCallSite.CB; 1385 1386 // Jump tables are only profitable if the retpoline mitigation is enabled. 1387 Attribute FSAttr = CB.getCaller()->getFnAttribute("target-features"); 1388 if (!FSAttr.isValid() || 1389 !FSAttr.getValueAsString().contains("+retpoline")) 1390 continue; 1391 1392 NumBranchFunnel++; 1393 if (RemarksEnabled) 1394 VCallSite.emitRemark("branch-funnel", 1395 JT->stripPointerCasts()->getName(), OREGetter); 1396 1397 // Pass the address of the vtable in the nest register, which is r10 on 1398 // x86_64. 1399 std::vector<Type *> NewArgs; 1400 NewArgs.push_back(Int8PtrTy); 1401 append_range(NewArgs, CB.getFunctionType()->params()); 1402 FunctionType *NewFT = 1403 FunctionType::get(CB.getFunctionType()->getReturnType(), NewArgs, 1404 CB.getFunctionType()->isVarArg()); 1405 PointerType *NewFTPtr = PointerType::getUnqual(NewFT); 1406 1407 IRBuilder<> IRB(&CB); 1408 std::vector<Value *> Args; 1409 Args.push_back(IRB.CreateBitCast(VCallSite.VTable, Int8PtrTy)); 1410 llvm::append_range(Args, CB.args()); 1411 1412 CallBase *NewCS = nullptr; 1413 if (isa<CallInst>(CB)) 1414 NewCS = IRB.CreateCall(NewFT, IRB.CreateBitCast(JT, NewFTPtr), Args); 1415 else 1416 NewCS = IRB.CreateInvoke(NewFT, IRB.CreateBitCast(JT, NewFTPtr), 1417 cast<InvokeInst>(CB).getNormalDest(), 1418 cast<InvokeInst>(CB).getUnwindDest(), Args); 1419 NewCS->setCallingConv(CB.getCallingConv()); 1420 1421 AttributeList Attrs = CB.getAttributes(); 1422 std::vector<AttributeSet> NewArgAttrs; 1423 NewArgAttrs.push_back(AttributeSet::get( 1424 M.getContext(), ArrayRef<Attribute>{Attribute::get( 1425 M.getContext(), Attribute::Nest)})); 1426 for (unsigned I = 0; I + 2 < Attrs.getNumAttrSets(); ++I) 1427 NewArgAttrs.push_back(Attrs.getParamAttrs(I)); 1428 NewCS->setAttributes( 1429 AttributeList::get(M.getContext(), Attrs.getFnAttrs(), 1430 Attrs.getRetAttrs(), NewArgAttrs)); 1431 1432 CB.replaceAllUsesWith(NewCS); 1433 CB.eraseFromParent(); 1434 1435 // This use is no longer unsafe. 1436 if (VCallSite.NumUnsafeUses) 1437 --*VCallSite.NumUnsafeUses; 1438 } 1439 // Don't mark as devirtualized because there may be callers compiled without 1440 // retpoline mitigation, which would mean that they are lowered to 1441 // llvm.type.test and therefore require an llvm.type.test resolution for the 1442 // type identifier. 1443 }; 1444 Apply(SlotInfo.CSInfo); 1445 for (auto &P : SlotInfo.ConstCSInfo) 1446 Apply(P.second); 1447 } 1448 1449 bool DevirtModule::tryEvaluateFunctionsWithArgs( 1450 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1451 ArrayRef<uint64_t> Args) { 1452 // Evaluate each function and store the result in each target's RetVal 1453 // field. 1454 for (VirtualCallTarget &Target : TargetsForSlot) { 1455 if (Target.Fn->arg_size() != Args.size() + 1) 1456 return false; 1457 1458 Evaluator Eval(M.getDataLayout(), nullptr); 1459 SmallVector<Constant *, 2> EvalArgs; 1460 EvalArgs.push_back( 1461 Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0))); 1462 for (unsigned I = 0; I != Args.size(); ++I) { 1463 auto *ArgTy = dyn_cast<IntegerType>( 1464 Target.Fn->getFunctionType()->getParamType(I + 1)); 1465 if (!ArgTy) 1466 return false; 1467 EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I])); 1468 } 1469 1470 Constant *RetVal; 1471 if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) || 1472 !isa<ConstantInt>(RetVal)) 1473 return false; 1474 Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue(); 1475 } 1476 return true; 1477 } 1478 1479 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1480 uint64_t TheRetVal) { 1481 for (auto Call : CSInfo.CallSites) { 1482 if (!OptimizedCalls.insert(&Call.CB).second) 1483 continue; 1484 NumUniformRetVal++; 1485 Call.replaceAndErase( 1486 "uniform-ret-val", FnName, RemarksEnabled, OREGetter, 1487 ConstantInt::get(cast<IntegerType>(Call.CB.getType()), TheRetVal)); 1488 } 1489 CSInfo.markDevirt(); 1490 } 1491 1492 bool DevirtModule::tryUniformRetValOpt( 1493 MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo, 1494 WholeProgramDevirtResolution::ByArg *Res) { 1495 // Uniform return value optimization. If all functions return the same 1496 // constant, replace all calls with that constant. 1497 uint64_t TheRetVal = TargetsForSlot[0].RetVal; 1498 for (const VirtualCallTarget &Target : TargetsForSlot) 1499 if (Target.RetVal != TheRetVal) 1500 return false; 1501 1502 if (CSInfo.isExported()) { 1503 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniformRetVal; 1504 Res->Info = TheRetVal; 1505 } 1506 1507 applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal); 1508 if (RemarksEnabled || AreStatisticsEnabled()) 1509 for (auto &&Target : TargetsForSlot) 1510 Target.WasDevirt = true; 1511 return true; 1512 } 1513 1514 std::string DevirtModule::getGlobalName(VTableSlot Slot, 1515 ArrayRef<uint64_t> Args, 1516 StringRef Name) { 1517 std::string FullName = "__typeid_"; 1518 raw_string_ostream OS(FullName); 1519 OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset; 1520 for (uint64_t Arg : Args) 1521 OS << '_' << Arg; 1522 OS << '_' << Name; 1523 return OS.str(); 1524 } 1525 1526 bool DevirtModule::shouldExportConstantsAsAbsoluteSymbols() { 1527 Triple T(M.getTargetTriple()); 1528 return T.isX86() && T.getObjectFormat() == Triple::ELF; 1529 } 1530 1531 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1532 StringRef Name, Constant *C) { 1533 GlobalAlias *GA = GlobalAlias::create(Int8Ty, 0, GlobalValue::ExternalLinkage, 1534 getGlobalName(Slot, Args, Name), C, &M); 1535 GA->setVisibility(GlobalValue::HiddenVisibility); 1536 } 1537 1538 void DevirtModule::exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1539 StringRef Name, uint32_t Const, 1540 uint32_t &Storage) { 1541 if (shouldExportConstantsAsAbsoluteSymbols()) { 1542 exportGlobal( 1543 Slot, Args, Name, 1544 ConstantExpr::getIntToPtr(ConstantInt::get(Int32Ty, Const), Int8PtrTy)); 1545 return; 1546 } 1547 1548 Storage = Const; 1549 } 1550 1551 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1552 StringRef Name) { 1553 Constant *C = 1554 M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Arr0Ty); 1555 auto *GV = dyn_cast<GlobalVariable>(C); 1556 if (GV) 1557 GV->setVisibility(GlobalValue::HiddenVisibility); 1558 return C; 1559 } 1560 1561 Constant *DevirtModule::importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1562 StringRef Name, IntegerType *IntTy, 1563 uint32_t Storage) { 1564 if (!shouldExportConstantsAsAbsoluteSymbols()) 1565 return ConstantInt::get(IntTy, Storage); 1566 1567 Constant *C = importGlobal(Slot, Args, Name); 1568 auto *GV = cast<GlobalVariable>(C->stripPointerCasts()); 1569 C = ConstantExpr::getPtrToInt(C, IntTy); 1570 1571 // We only need to set metadata if the global is newly created, in which 1572 // case it would not have hidden visibility. 1573 if (GV->hasMetadata(LLVMContext::MD_absolute_symbol)) 1574 return C; 1575 1576 auto SetAbsRange = [&](uint64_t Min, uint64_t Max) { 1577 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min)); 1578 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max)); 1579 GV->setMetadata(LLVMContext::MD_absolute_symbol, 1580 MDNode::get(M.getContext(), {MinC, MaxC})); 1581 }; 1582 unsigned AbsWidth = IntTy->getBitWidth(); 1583 if (AbsWidth == IntPtrTy->getBitWidth()) 1584 SetAbsRange(~0ull, ~0ull); // Full set. 1585 else 1586 SetAbsRange(0, 1ull << AbsWidth); 1587 return C; 1588 } 1589 1590 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1591 bool IsOne, 1592 Constant *UniqueMemberAddr) { 1593 for (auto &&Call : CSInfo.CallSites) { 1594 if (!OptimizedCalls.insert(&Call.CB).second) 1595 continue; 1596 IRBuilder<> B(&Call.CB); 1597 Value *Cmp = 1598 B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE, Call.VTable, 1599 B.CreateBitCast(UniqueMemberAddr, Call.VTable->getType())); 1600 Cmp = B.CreateZExt(Cmp, Call.CB.getType()); 1601 NumUniqueRetVal++; 1602 Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter, 1603 Cmp); 1604 } 1605 CSInfo.markDevirt(); 1606 } 1607 1608 Constant *DevirtModule::getMemberAddr(const TypeMemberInfo *M) { 1609 Constant *C = ConstantExpr::getBitCast(M->Bits->GV, Int8PtrTy); 1610 return ConstantExpr::getGetElementPtr(Int8Ty, C, 1611 ConstantInt::get(Int64Ty, M->Offset)); 1612 } 1613 1614 bool DevirtModule::tryUniqueRetValOpt( 1615 unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1616 CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res, 1617 VTableSlot Slot, ArrayRef<uint64_t> Args) { 1618 // IsOne controls whether we look for a 0 or a 1. 1619 auto tryUniqueRetValOptFor = [&](bool IsOne) { 1620 const TypeMemberInfo *UniqueMember = nullptr; 1621 for (const VirtualCallTarget &Target : TargetsForSlot) { 1622 if (Target.RetVal == (IsOne ? 1 : 0)) { 1623 if (UniqueMember) 1624 return false; 1625 UniqueMember = Target.TM; 1626 } 1627 } 1628 1629 // We should have found a unique member or bailed out by now. We already 1630 // checked for a uniform return value in tryUniformRetValOpt. 1631 assert(UniqueMember); 1632 1633 Constant *UniqueMemberAddr = getMemberAddr(UniqueMember); 1634 if (CSInfo.isExported()) { 1635 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniqueRetVal; 1636 Res->Info = IsOne; 1637 1638 exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr); 1639 } 1640 1641 // Replace each call with the comparison. 1642 applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne, 1643 UniqueMemberAddr); 1644 1645 // Update devirtualization statistics for targets. 1646 if (RemarksEnabled || AreStatisticsEnabled()) 1647 for (auto &&Target : TargetsForSlot) 1648 Target.WasDevirt = true; 1649 1650 return true; 1651 }; 1652 1653 if (BitWidth == 1) { 1654 if (tryUniqueRetValOptFor(true)) 1655 return true; 1656 if (tryUniqueRetValOptFor(false)) 1657 return true; 1658 } 1659 return false; 1660 } 1661 1662 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 1663 Constant *Byte, Constant *Bit) { 1664 for (auto Call : CSInfo.CallSites) { 1665 if (!OptimizedCalls.insert(&Call.CB).second) 1666 continue; 1667 auto *RetType = cast<IntegerType>(Call.CB.getType()); 1668 IRBuilder<> B(&Call.CB); 1669 Value *Addr = 1670 B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte); 1671 if (RetType->getBitWidth() == 1) { 1672 Value *Bits = B.CreateLoad(Int8Ty, Addr); 1673 Value *BitsAndBit = B.CreateAnd(Bits, Bit); 1674 auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0)); 1675 NumVirtConstProp1Bit++; 1676 Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled, 1677 OREGetter, IsBitSet); 1678 } else { 1679 Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo()); 1680 Value *Val = B.CreateLoad(RetType, ValAddr); 1681 NumVirtConstProp++; 1682 Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled, 1683 OREGetter, Val); 1684 } 1685 } 1686 CSInfo.markDevirt(); 1687 } 1688 1689 bool DevirtModule::tryVirtualConstProp( 1690 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1691 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1692 // This only works if the function returns an integer. 1693 auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType()); 1694 if (!RetType) 1695 return false; 1696 unsigned BitWidth = RetType->getBitWidth(); 1697 if (BitWidth > 64) 1698 return false; 1699 1700 // Make sure that each function is defined, does not access memory, takes at 1701 // least one argument, does not use its first argument (which we assume is 1702 // 'this'), and has the same return type. 1703 // 1704 // Note that we test whether this copy of the function is readnone, rather 1705 // than testing function attributes, which must hold for any copy of the 1706 // function, even a less optimized version substituted at link time. This is 1707 // sound because the virtual constant propagation optimizations effectively 1708 // inline all implementations of the virtual function into each call site, 1709 // rather than using function attributes to perform local optimization. 1710 for (VirtualCallTarget &Target : TargetsForSlot) { 1711 if (Target.Fn->isDeclaration() || 1712 computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) != 1713 FMRB_DoesNotAccessMemory || 1714 Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() || 1715 Target.Fn->getReturnType() != RetType) 1716 return false; 1717 } 1718 1719 for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) { 1720 if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first)) 1721 continue; 1722 1723 WholeProgramDevirtResolution::ByArg *ResByArg = nullptr; 1724 if (Res) 1725 ResByArg = &Res->ResByArg[CSByConstantArg.first]; 1726 1727 if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg)) 1728 continue; 1729 1730 if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second, 1731 ResByArg, Slot, CSByConstantArg.first)) 1732 continue; 1733 1734 // Find an allocation offset in bits in all vtables associated with the 1735 // type. 1736 uint64_t AllocBefore = 1737 findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth); 1738 uint64_t AllocAfter = 1739 findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth); 1740 1741 // Calculate the total amount of padding needed to store a value at both 1742 // ends of the object. 1743 uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0; 1744 for (auto &&Target : TargetsForSlot) { 1745 TotalPaddingBefore += std::max<int64_t>( 1746 (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0); 1747 TotalPaddingAfter += std::max<int64_t>( 1748 (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0); 1749 } 1750 1751 // If the amount of padding is too large, give up. 1752 // FIXME: do something smarter here. 1753 if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128) 1754 continue; 1755 1756 // Calculate the offset to the value as a (possibly negative) byte offset 1757 // and (if applicable) a bit offset, and store the values in the targets. 1758 int64_t OffsetByte; 1759 uint64_t OffsetBit; 1760 if (TotalPaddingBefore <= TotalPaddingAfter) 1761 setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte, 1762 OffsetBit); 1763 else 1764 setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte, 1765 OffsetBit); 1766 1767 if (RemarksEnabled || AreStatisticsEnabled()) 1768 for (auto &&Target : TargetsForSlot) 1769 Target.WasDevirt = true; 1770 1771 1772 if (CSByConstantArg.second.isExported()) { 1773 ResByArg->TheKind = WholeProgramDevirtResolution::ByArg::VirtualConstProp; 1774 exportConstant(Slot, CSByConstantArg.first, "byte", OffsetByte, 1775 ResByArg->Byte); 1776 exportConstant(Slot, CSByConstantArg.first, "bit", 1ULL << OffsetBit, 1777 ResByArg->Bit); 1778 } 1779 1780 // Rewrite each call to a load from OffsetByte/OffsetBit. 1781 Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte); 1782 Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit); 1783 applyVirtualConstProp(CSByConstantArg.second, 1784 TargetsForSlot[0].Fn->getName(), ByteConst, BitConst); 1785 } 1786 return true; 1787 } 1788 1789 void DevirtModule::rebuildGlobal(VTableBits &B) { 1790 if (B.Before.Bytes.empty() && B.After.Bytes.empty()) 1791 return; 1792 1793 // Align the before byte array to the global's minimum alignment so that we 1794 // don't break any alignment requirements on the global. 1795 Align Alignment = M.getDataLayout().getValueOrABITypeAlignment( 1796 B.GV->getAlign(), B.GV->getValueType()); 1797 B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), Alignment)); 1798 1799 // Before was stored in reverse order; flip it now. 1800 for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I) 1801 std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]); 1802 1803 // Build an anonymous global containing the before bytes, followed by the 1804 // original initializer, followed by the after bytes. 1805 auto NewInit = ConstantStruct::getAnon( 1806 {ConstantDataArray::get(M.getContext(), B.Before.Bytes), 1807 B.GV->getInitializer(), 1808 ConstantDataArray::get(M.getContext(), B.After.Bytes)}); 1809 auto NewGV = 1810 new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(), 1811 GlobalVariable::PrivateLinkage, NewInit, "", B.GV); 1812 NewGV->setSection(B.GV->getSection()); 1813 NewGV->setComdat(B.GV->getComdat()); 1814 NewGV->setAlignment(B.GV->getAlign()); 1815 1816 // Copy the original vtable's metadata to the anonymous global, adjusting 1817 // offsets as required. 1818 NewGV->copyMetadata(B.GV, B.Before.Bytes.size()); 1819 1820 // Build an alias named after the original global, pointing at the second 1821 // element (the original initializer). 1822 auto Alias = GlobalAlias::create( 1823 B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "", 1824 ConstantExpr::getGetElementPtr( 1825 NewInit->getType(), NewGV, 1826 ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0), 1827 ConstantInt::get(Int32Ty, 1)}), 1828 &M); 1829 Alias->setVisibility(B.GV->getVisibility()); 1830 Alias->takeName(B.GV); 1831 1832 B.GV->replaceAllUsesWith(Alias); 1833 B.GV->eraseFromParent(); 1834 } 1835 1836 bool DevirtModule::areRemarksEnabled() { 1837 const auto &FL = M.getFunctionList(); 1838 for (const Function &Fn : FL) { 1839 const auto &BBL = Fn.getBasicBlockList(); 1840 if (BBL.empty()) 1841 continue; 1842 auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front()); 1843 return DI.isEnabled(); 1844 } 1845 return false; 1846 } 1847 1848 void DevirtModule::scanTypeTestUsers( 1849 Function *TypeTestFunc, 1850 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 1851 // Find all virtual calls via a virtual table pointer %p under an assumption 1852 // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p 1853 // points to a member of the type identifier %md. Group calls by (type ID, 1854 // offset) pair (effectively the identity of the virtual function) and store 1855 // to CallSlots. 1856 for (Use &U : llvm::make_early_inc_range(TypeTestFunc->uses())) { 1857 auto *CI = dyn_cast<CallInst>(U.getUser()); 1858 if (!CI) 1859 continue; 1860 1861 // Search for virtual calls based on %p and add them to DevirtCalls. 1862 SmallVector<DevirtCallSite, 1> DevirtCalls; 1863 SmallVector<CallInst *, 1> Assumes; 1864 auto &DT = LookupDomTree(*CI->getFunction()); 1865 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT); 1866 1867 Metadata *TypeId = 1868 cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata(); 1869 // If we found any, add them to CallSlots. 1870 if (!Assumes.empty()) { 1871 Value *Ptr = CI->getArgOperand(0)->stripPointerCasts(); 1872 for (DevirtCallSite Call : DevirtCalls) 1873 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, nullptr); 1874 } 1875 1876 auto RemoveTypeTestAssumes = [&]() { 1877 // We no longer need the assumes or the type test. 1878 for (auto Assume : Assumes) 1879 Assume->eraseFromParent(); 1880 // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we 1881 // may use the vtable argument later. 1882 if (CI->use_empty()) 1883 CI->eraseFromParent(); 1884 }; 1885 1886 // At this point we could remove all type test assume sequences, as they 1887 // were originally inserted for WPD. However, we can keep these in the 1888 // code stream for later analysis (e.g. to help drive more efficient ICP 1889 // sequences). They will eventually be removed by a second LowerTypeTests 1890 // invocation that cleans them up. In order to do this correctly, the first 1891 // LowerTypeTests invocation needs to know that they have "Unknown" type 1892 // test resolution, so that they aren't treated as Unsat and lowered to 1893 // False, which will break any uses on assumes. Below we remove any type 1894 // test assumes that will not be treated as Unknown by LTT. 1895 1896 // The type test assumes will be treated by LTT as Unsat if the type id is 1897 // not used on a global (in which case it has no entry in the TypeIdMap). 1898 if (!TypeIdMap.count(TypeId)) 1899 RemoveTypeTestAssumes(); 1900 1901 // For ThinLTO importing, we need to remove the type test assumes if this is 1902 // an MDString type id without a corresponding TypeIdSummary. Any 1903 // non-MDString type ids are ignored and treated as Unknown by LTT, so their 1904 // type test assumes can be kept. If the MDString type id is missing a 1905 // TypeIdSummary (e.g. because there was no use on a vcall, preventing the 1906 // exporting phase of WPD from analyzing it), then it would be treated as 1907 // Unsat by LTT and we need to remove its type test assumes here. If not 1908 // used on a vcall we don't need them for later optimization use in any 1909 // case. 1910 else if (ImportSummary && isa<MDString>(TypeId)) { 1911 const TypeIdSummary *TidSummary = 1912 ImportSummary->getTypeIdSummary(cast<MDString>(TypeId)->getString()); 1913 if (!TidSummary) 1914 RemoveTypeTestAssumes(); 1915 else 1916 // If one was created it should not be Unsat, because if we reached here 1917 // the type id was used on a global. 1918 assert(TidSummary->TTRes.TheKind != TypeTestResolution::Unsat); 1919 } 1920 } 1921 } 1922 1923 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) { 1924 Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test); 1925 1926 for (Use &U : llvm::make_early_inc_range(TypeCheckedLoadFunc->uses())) { 1927 auto *CI = dyn_cast<CallInst>(U.getUser()); 1928 if (!CI) 1929 continue; 1930 1931 Value *Ptr = CI->getArgOperand(0); 1932 Value *Offset = CI->getArgOperand(1); 1933 Value *TypeIdValue = CI->getArgOperand(2); 1934 Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata(); 1935 1936 SmallVector<DevirtCallSite, 1> DevirtCalls; 1937 SmallVector<Instruction *, 1> LoadedPtrs; 1938 SmallVector<Instruction *, 1> Preds; 1939 bool HasNonCallUses = false; 1940 auto &DT = LookupDomTree(*CI->getFunction()); 1941 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds, 1942 HasNonCallUses, CI, DT); 1943 1944 // Start by generating "pessimistic" code that explicitly loads the function 1945 // pointer from the vtable and performs the type check. If possible, we will 1946 // eliminate the load and the type check later. 1947 1948 // If possible, only generate the load at the point where it is used. 1949 // This helps avoid unnecessary spills. 1950 IRBuilder<> LoadB( 1951 (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI); 1952 Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset); 1953 Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy)); 1954 Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr); 1955 1956 for (Instruction *LoadedPtr : LoadedPtrs) { 1957 LoadedPtr->replaceAllUsesWith(LoadedValue); 1958 LoadedPtr->eraseFromParent(); 1959 } 1960 1961 // Likewise for the type test. 1962 IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI); 1963 CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue}); 1964 1965 for (Instruction *Pred : Preds) { 1966 Pred->replaceAllUsesWith(TypeTestCall); 1967 Pred->eraseFromParent(); 1968 } 1969 1970 // We have already erased any extractvalue instructions that refer to the 1971 // intrinsic call, but the intrinsic may have other non-extractvalue uses 1972 // (although this is unlikely). In that case, explicitly build a pair and 1973 // RAUW it. 1974 if (!CI->use_empty()) { 1975 Value *Pair = PoisonValue::get(CI->getType()); 1976 IRBuilder<> B(CI); 1977 Pair = B.CreateInsertValue(Pair, LoadedValue, {0}); 1978 Pair = B.CreateInsertValue(Pair, TypeTestCall, {1}); 1979 CI->replaceAllUsesWith(Pair); 1980 } 1981 1982 // The number of unsafe uses is initially the number of uses. 1983 auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall]; 1984 NumUnsafeUses = DevirtCalls.size(); 1985 1986 // If the function pointer has a non-call user, we cannot eliminate the type 1987 // check, as one of those users may eventually call the pointer. Increment 1988 // the unsafe use count to make sure it cannot reach zero. 1989 if (HasNonCallUses) 1990 ++NumUnsafeUses; 1991 for (DevirtCallSite Call : DevirtCalls) { 1992 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, 1993 &NumUnsafeUses); 1994 } 1995 1996 CI->eraseFromParent(); 1997 } 1998 } 1999 2000 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) { 2001 auto *TypeId = dyn_cast<MDString>(Slot.TypeID); 2002 if (!TypeId) 2003 return; 2004 const TypeIdSummary *TidSummary = 2005 ImportSummary->getTypeIdSummary(TypeId->getString()); 2006 if (!TidSummary) 2007 return; 2008 auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset); 2009 if (ResI == TidSummary->WPDRes.end()) 2010 return; 2011 const WholeProgramDevirtResolution &Res = ResI->second; 2012 2013 if (Res.TheKind == WholeProgramDevirtResolution::SingleImpl) { 2014 assert(!Res.SingleImplName.empty()); 2015 // The type of the function in the declaration is irrelevant because every 2016 // call site will cast it to the correct type. 2017 Constant *SingleImpl = 2018 cast<Constant>(M.getOrInsertFunction(Res.SingleImplName, 2019 Type::getVoidTy(M.getContext())) 2020 .getCallee()); 2021 2022 // This is the import phase so we should not be exporting anything. 2023 bool IsExported = false; 2024 applySingleImplDevirt(SlotInfo, SingleImpl, IsExported); 2025 assert(!IsExported); 2026 } 2027 2028 for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) { 2029 auto I = Res.ResByArg.find(CSByConstantArg.first); 2030 if (I == Res.ResByArg.end()) 2031 continue; 2032 auto &ResByArg = I->second; 2033 // FIXME: We should figure out what to do about the "function name" argument 2034 // to the apply* functions, as the function names are unavailable during the 2035 // importing phase. For now we just pass the empty string. This does not 2036 // impact correctness because the function names are just used for remarks. 2037 switch (ResByArg.TheKind) { 2038 case WholeProgramDevirtResolution::ByArg::UniformRetVal: 2039 applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info); 2040 break; 2041 case WholeProgramDevirtResolution::ByArg::UniqueRetVal: { 2042 Constant *UniqueMemberAddr = 2043 importGlobal(Slot, CSByConstantArg.first, "unique_member"); 2044 applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info, 2045 UniqueMemberAddr); 2046 break; 2047 } 2048 case WholeProgramDevirtResolution::ByArg::VirtualConstProp: { 2049 Constant *Byte = importConstant(Slot, CSByConstantArg.first, "byte", 2050 Int32Ty, ResByArg.Byte); 2051 Constant *Bit = importConstant(Slot, CSByConstantArg.first, "bit", Int8Ty, 2052 ResByArg.Bit); 2053 applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit); 2054 break; 2055 } 2056 default: 2057 break; 2058 } 2059 } 2060 2061 if (Res.TheKind == WholeProgramDevirtResolution::BranchFunnel) { 2062 // The type of the function is irrelevant, because it's bitcast at calls 2063 // anyhow. 2064 Constant *JT = cast<Constant>( 2065 M.getOrInsertFunction(getGlobalName(Slot, {}, "branch_funnel"), 2066 Type::getVoidTy(M.getContext())) 2067 .getCallee()); 2068 bool IsExported = false; 2069 applyICallBranchFunnel(SlotInfo, JT, IsExported); 2070 assert(!IsExported); 2071 } 2072 } 2073 2074 void DevirtModule::removeRedundantTypeTests() { 2075 auto True = ConstantInt::getTrue(M.getContext()); 2076 for (auto &&U : NumUnsafeUsesForTypeTest) { 2077 if (U.second == 0) { 2078 U.first->replaceAllUsesWith(True); 2079 U.first->eraseFromParent(); 2080 } 2081 } 2082 } 2083 2084 ValueInfo 2085 DevirtModule::lookUpFunctionValueInfo(Function *TheFn, 2086 ModuleSummaryIndex *ExportSummary) { 2087 assert((ExportSummary != nullptr) && 2088 "Caller guarantees ExportSummary is not nullptr"); 2089 2090 const auto TheFnGUID = TheFn->getGUID(); 2091 const auto TheFnGUIDWithExportedName = GlobalValue::getGUID(TheFn->getName()); 2092 // Look up ValueInfo with the GUID in the current linkage. 2093 ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFnGUID); 2094 // If no entry is found and GUID is different from GUID computed using 2095 // exported name, look up ValueInfo with the exported name unconditionally. 2096 // This is a fallback. 2097 // 2098 // The reason to have a fallback: 2099 // 1. LTO could enable global value internalization via 2100 // `enable-lto-internalization`. 2101 // 2. The GUID in ExportedSummary is computed using exported name. 2102 if ((!TheFnVI) && (TheFnGUID != TheFnGUIDWithExportedName)) { 2103 TheFnVI = ExportSummary->getValueInfo(TheFnGUIDWithExportedName); 2104 } 2105 return TheFnVI; 2106 } 2107 2108 bool DevirtModule::mustBeUnreachableFunction( 2109 Function *const F, ModuleSummaryIndex *ExportSummary) { 2110 // First, learn unreachability by analyzing function IR. 2111 if (!F->isDeclaration()) { 2112 // A function must be unreachable if its entry block ends with an 2113 // 'unreachable'. 2114 return isa<UnreachableInst>(F->getEntryBlock().getTerminator()); 2115 } 2116 // Learn unreachability from ExportSummary if ExportSummary is present. 2117 return ExportSummary && 2118 ::mustBeUnreachableFunction( 2119 DevirtModule::lookUpFunctionValueInfo(F, ExportSummary)); 2120 } 2121 2122 bool DevirtModule::run() { 2123 // If only some of the modules were split, we cannot correctly perform 2124 // this transformation. We already checked for the presense of type tests 2125 // with partially split modules during the thin link, and would have emitted 2126 // an error if any were found, so here we can simply return. 2127 if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) || 2128 (ImportSummary && ImportSummary->partiallySplitLTOUnits())) 2129 return false; 2130 2131 Function *TypeTestFunc = 2132 M.getFunction(Intrinsic::getName(Intrinsic::type_test)); 2133 Function *TypeCheckedLoadFunc = 2134 M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load)); 2135 Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume)); 2136 2137 // Normally if there are no users of the devirtualization intrinsics in the 2138 // module, this pass has nothing to do. But if we are exporting, we also need 2139 // to handle any users that appear only in the function summaries. 2140 if (!ExportSummary && 2141 (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc || 2142 AssumeFunc->use_empty()) && 2143 (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty())) 2144 return false; 2145 2146 // Rebuild type metadata into a map for easy lookup. 2147 std::vector<VTableBits> Bits; 2148 DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap; 2149 buildTypeIdentifierMap(Bits, TypeIdMap); 2150 2151 if (TypeTestFunc && AssumeFunc) 2152 scanTypeTestUsers(TypeTestFunc, TypeIdMap); 2153 2154 if (TypeCheckedLoadFunc) 2155 scanTypeCheckedLoadUsers(TypeCheckedLoadFunc); 2156 2157 if (ImportSummary) { 2158 for (auto &S : CallSlots) 2159 importResolution(S.first, S.second); 2160 2161 removeRedundantTypeTests(); 2162 2163 // We have lowered or deleted the type intrinsics, so we will no longer have 2164 // enough information to reason about the liveness of virtual function 2165 // pointers in GlobalDCE. 2166 for (GlobalVariable &GV : M.globals()) 2167 GV.eraseMetadata(LLVMContext::MD_vcall_visibility); 2168 2169 // The rest of the code is only necessary when exporting or during regular 2170 // LTO, so we are done. 2171 return true; 2172 } 2173 2174 if (TypeIdMap.empty()) 2175 return true; 2176 2177 // Collect information from summary about which calls to try to devirtualize. 2178 if (ExportSummary) { 2179 DenseMap<GlobalValue::GUID, TinyPtrVector<Metadata *>> MetadataByGUID; 2180 for (auto &P : TypeIdMap) { 2181 if (auto *TypeId = dyn_cast<MDString>(P.first)) 2182 MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back( 2183 TypeId); 2184 } 2185 2186 for (auto &P : *ExportSummary) { 2187 for (auto &S : P.second.SummaryList) { 2188 auto *FS = dyn_cast<FunctionSummary>(S.get()); 2189 if (!FS) 2190 continue; 2191 // FIXME: Only add live functions. 2192 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 2193 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 2194 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 2195 } 2196 } 2197 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 2198 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 2199 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 2200 } 2201 } 2202 for (const FunctionSummary::ConstVCall &VC : 2203 FS->type_test_assume_const_vcalls()) { 2204 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 2205 CallSlots[{MD, VC.VFunc.Offset}] 2206 .ConstCSInfo[VC.Args] 2207 .addSummaryTypeTestAssumeUser(FS); 2208 } 2209 } 2210 for (const FunctionSummary::ConstVCall &VC : 2211 FS->type_checked_load_const_vcalls()) { 2212 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 2213 CallSlots[{MD, VC.VFunc.Offset}] 2214 .ConstCSInfo[VC.Args] 2215 .addSummaryTypeCheckedLoadUser(FS); 2216 } 2217 } 2218 } 2219 } 2220 } 2221 2222 // For each (type, offset) pair: 2223 bool DidVirtualConstProp = false; 2224 std::map<std::string, Function*> DevirtTargets; 2225 for (auto &S : CallSlots) { 2226 // Search each of the members of the type identifier for the virtual 2227 // function implementation at offset S.first.ByteOffset, and add to 2228 // TargetsForSlot. 2229 std::vector<VirtualCallTarget> TargetsForSlot; 2230 WholeProgramDevirtResolution *Res = nullptr; 2231 const std::set<TypeMemberInfo> &TypeMemberInfos = TypeIdMap[S.first.TypeID]; 2232 if (ExportSummary && isa<MDString>(S.first.TypeID) && 2233 TypeMemberInfos.size()) 2234 // For any type id used on a global's type metadata, create the type id 2235 // summary resolution regardless of whether we can devirtualize, so that 2236 // lower type tests knows the type id is not Unsat. If it was not used on 2237 // a global's type metadata, the TypeIdMap entry set will be empty, and 2238 // we don't want to create an entry (with the default Unknown type 2239 // resolution), which can prevent detection of the Unsat. 2240 Res = &ExportSummary 2241 ->getOrInsertTypeIdSummary( 2242 cast<MDString>(S.first.TypeID)->getString()) 2243 .WPDRes[S.first.ByteOffset]; 2244 if (tryFindVirtualCallTargets(TargetsForSlot, TypeMemberInfos, 2245 S.first.ByteOffset, ExportSummary)) { 2246 2247 if (!trySingleImplDevirt(ExportSummary, TargetsForSlot, S.second, Res)) { 2248 DidVirtualConstProp |= 2249 tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first); 2250 2251 tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first); 2252 } 2253 2254 // Collect functions devirtualized at least for one call site for stats. 2255 if (RemarksEnabled || AreStatisticsEnabled()) 2256 for (const auto &T : TargetsForSlot) 2257 if (T.WasDevirt) 2258 DevirtTargets[std::string(T.Fn->getName())] = T.Fn; 2259 } 2260 2261 // CFI-specific: if we are exporting and any llvm.type.checked.load 2262 // intrinsics were *not* devirtualized, we need to add the resulting 2263 // llvm.type.test intrinsics to the function summaries so that the 2264 // LowerTypeTests pass will export them. 2265 if (ExportSummary && isa<MDString>(S.first.TypeID)) { 2266 auto GUID = 2267 GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString()); 2268 for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers) 2269 FS->addTypeTest(GUID); 2270 for (auto &CCS : S.second.ConstCSInfo) 2271 for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers) 2272 FS->addTypeTest(GUID); 2273 } 2274 } 2275 2276 if (RemarksEnabled) { 2277 // Generate remarks for each devirtualized function. 2278 for (const auto &DT : DevirtTargets) { 2279 Function *F = DT.second; 2280 2281 using namespace ore; 2282 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F) 2283 << "devirtualized " 2284 << NV("FunctionName", DT.first)); 2285 } 2286 } 2287 2288 NumDevirtTargets += DevirtTargets.size(); 2289 2290 removeRedundantTypeTests(); 2291 2292 // Rebuild each global we touched as part of virtual constant propagation to 2293 // include the before and after bytes. 2294 if (DidVirtualConstProp) 2295 for (VTableBits &B : Bits) 2296 rebuildGlobal(B); 2297 2298 // We have lowered or deleted the type intrinsics, so we will no longer have 2299 // enough information to reason about the liveness of virtual function 2300 // pointers in GlobalDCE. 2301 for (GlobalVariable &GV : M.globals()) 2302 GV.eraseMetadata(LLVMContext::MD_vcall_visibility); 2303 2304 return true; 2305 } 2306 2307 void DevirtIndex::run() { 2308 if (ExportSummary.typeIdCompatibleVtableMap().empty()) 2309 return; 2310 2311 DenseMap<GlobalValue::GUID, std::vector<StringRef>> NameByGUID; 2312 for (auto &P : ExportSummary.typeIdCompatibleVtableMap()) { 2313 NameByGUID[GlobalValue::getGUID(P.first)].push_back(P.first); 2314 } 2315 2316 // Collect information from summary about which calls to try to devirtualize. 2317 for (auto &P : ExportSummary) { 2318 for (auto &S : P.second.SummaryList) { 2319 auto *FS = dyn_cast<FunctionSummary>(S.get()); 2320 if (!FS) 2321 continue; 2322 // FIXME: Only add live functions. 2323 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 2324 for (StringRef Name : NameByGUID[VF.GUID]) { 2325 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 2326 } 2327 } 2328 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 2329 for (StringRef Name : NameByGUID[VF.GUID]) { 2330 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 2331 } 2332 } 2333 for (const FunctionSummary::ConstVCall &VC : 2334 FS->type_test_assume_const_vcalls()) { 2335 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 2336 CallSlots[{Name, VC.VFunc.Offset}] 2337 .ConstCSInfo[VC.Args] 2338 .addSummaryTypeTestAssumeUser(FS); 2339 } 2340 } 2341 for (const FunctionSummary::ConstVCall &VC : 2342 FS->type_checked_load_const_vcalls()) { 2343 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 2344 CallSlots[{Name, VC.VFunc.Offset}] 2345 .ConstCSInfo[VC.Args] 2346 .addSummaryTypeCheckedLoadUser(FS); 2347 } 2348 } 2349 } 2350 } 2351 2352 std::set<ValueInfo> DevirtTargets; 2353 // For each (type, offset) pair: 2354 for (auto &S : CallSlots) { 2355 // Search each of the members of the type identifier for the virtual 2356 // function implementation at offset S.first.ByteOffset, and add to 2357 // TargetsForSlot. 2358 std::vector<ValueInfo> TargetsForSlot; 2359 auto TidSummary = ExportSummary.getTypeIdCompatibleVtableSummary(S.first.TypeID); 2360 assert(TidSummary); 2361 // Create the type id summary resolution regardlness of whether we can 2362 // devirtualize, so that lower type tests knows the type id is used on 2363 // a global and not Unsat. 2364 WholeProgramDevirtResolution *Res = 2365 &ExportSummary.getOrInsertTypeIdSummary(S.first.TypeID) 2366 .WPDRes[S.first.ByteOffset]; 2367 if (tryFindVirtualCallTargets(TargetsForSlot, *TidSummary, 2368 S.first.ByteOffset)) { 2369 2370 if (!trySingleImplDevirt(TargetsForSlot, S.first, S.second, Res, 2371 DevirtTargets)) 2372 continue; 2373 } 2374 } 2375 2376 // Optionally have the thin link print message for each devirtualized 2377 // function. 2378 if (PrintSummaryDevirt) 2379 for (const auto &DT : DevirtTargets) 2380 errs() << "Devirtualized call to " << DT << "\n"; 2381 2382 NumDevirtTargets += DevirtTargets.size(); 2383 } 2384