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::init(false), cl::Hidden, 167 cl::ZeroOrMore, 168 cl::desc("Disable whole program visibility (overrides enabling options)")); 169 170 /// Provide way to prevent certain function from being devirtualized 171 static cl::list<std::string> 172 SkipFunctionNames("wholeprogramdevirt-skip", 173 cl::desc("Prevent function(s) from being devirtualized"), 174 cl::Hidden, cl::ZeroOrMore, cl::CommaSeparated); 175 176 /// Mechanism to add runtime checking of devirtualization decisions, optionally 177 /// trapping or falling back to indirect call on any that are not correct. 178 /// Trapping mode is useful for debugging undefined behavior leading to failures 179 /// with WPD. Fallback mode is useful for ensuring safety when whole program 180 /// visibility may be compromised. 181 enum WPDCheckMode { None, Trap, Fallback }; 182 static cl::opt<WPDCheckMode> DevirtCheckMode( 183 "wholeprogramdevirt-check", cl::Hidden, 184 cl::desc("Type of checking for incorrect devirtualizations"), 185 cl::values(clEnumValN(WPDCheckMode::None, "none", "No checking"), 186 clEnumValN(WPDCheckMode::Trap, "trap", "Trap when incorrect"), 187 clEnumValN(WPDCheckMode::Fallback, "fallback", 188 "Fallback to indirect when incorrect"))); 189 190 namespace { 191 struct PatternList { 192 std::vector<GlobPattern> Patterns; 193 template <class T> void init(const T &StringList) { 194 for (const auto &S : StringList) 195 if (Expected<GlobPattern> Pat = GlobPattern::create(S)) 196 Patterns.push_back(std::move(*Pat)); 197 } 198 bool match(StringRef S) { 199 for (const GlobPattern &P : Patterns) 200 if (P.match(S)) 201 return true; 202 return false; 203 } 204 }; 205 } // namespace 206 207 // Find the minimum offset that we may store a value of size Size bits at. If 208 // IsAfter is set, look for an offset before the object, otherwise look for an 209 // offset after the object. 210 uint64_t 211 wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets, 212 bool IsAfter, uint64_t Size) { 213 // Find a minimum offset taking into account only vtable sizes. 214 uint64_t MinByte = 0; 215 for (const VirtualCallTarget &Target : Targets) { 216 if (IsAfter) 217 MinByte = std::max(MinByte, Target.minAfterBytes()); 218 else 219 MinByte = std::max(MinByte, Target.minBeforeBytes()); 220 } 221 222 // Build a vector of arrays of bytes covering, for each target, a slice of the 223 // used region (see AccumBitVector::BytesUsed in 224 // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively, 225 // this aligns the used regions to start at MinByte. 226 // 227 // In this example, A, B and C are vtables, # is a byte already allocated for 228 // a virtual function pointer, AAAA... (etc.) are the used regions for the 229 // vtables and Offset(X) is the value computed for the Offset variable below 230 // for X. 231 // 232 // Offset(A) 233 // | | 234 // |MinByte 235 // A: ################AAAAAAAA|AAAAAAAA 236 // B: ########BBBBBBBBBBBBBBBB|BBBB 237 // C: ########################|CCCCCCCCCCCCCCCC 238 // | Offset(B) | 239 // 240 // This code produces the slices of A, B and C that appear after the divider 241 // at MinByte. 242 std::vector<ArrayRef<uint8_t>> Used; 243 for (const VirtualCallTarget &Target : Targets) { 244 ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed 245 : Target.TM->Bits->Before.BytesUsed; 246 uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes() 247 : MinByte - Target.minBeforeBytes(); 248 249 // Disregard used regions that are smaller than Offset. These are 250 // effectively all-free regions that do not need to be checked. 251 if (VTUsed.size() > Offset) 252 Used.push_back(VTUsed.slice(Offset)); 253 } 254 255 if (Size == 1) { 256 // Find a free bit in each member of Used. 257 for (unsigned I = 0;; ++I) { 258 uint8_t BitsUsed = 0; 259 for (auto &&B : Used) 260 if (I < B.size()) 261 BitsUsed |= B[I]; 262 if (BitsUsed != 0xff) 263 return (MinByte + I) * 8 + 264 countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined); 265 } 266 } else { 267 // Find a free (Size/8) byte region in each member of Used. 268 // FIXME: see if alignment helps. 269 for (unsigned I = 0;; ++I) { 270 for (auto &&B : Used) { 271 unsigned Byte = 0; 272 while ((I + Byte) < B.size() && Byte < (Size / 8)) { 273 if (B[I + Byte]) 274 goto NextI; 275 ++Byte; 276 } 277 } 278 return (MinByte + I) * 8; 279 NextI:; 280 } 281 } 282 } 283 284 void wholeprogramdevirt::setBeforeReturnValues( 285 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore, 286 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 287 if (BitWidth == 1) 288 OffsetByte = -(AllocBefore / 8 + 1); 289 else 290 OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8); 291 OffsetBit = AllocBefore % 8; 292 293 for (VirtualCallTarget &Target : Targets) { 294 if (BitWidth == 1) 295 Target.setBeforeBit(AllocBefore); 296 else 297 Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8); 298 } 299 } 300 301 void wholeprogramdevirt::setAfterReturnValues( 302 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter, 303 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 304 if (BitWidth == 1) 305 OffsetByte = AllocAfter / 8; 306 else 307 OffsetByte = (AllocAfter + 7) / 8; 308 OffsetBit = AllocAfter % 8; 309 310 for (VirtualCallTarget &Target : Targets) { 311 if (BitWidth == 1) 312 Target.setAfterBit(AllocAfter); 313 else 314 Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8); 315 } 316 } 317 318 VirtualCallTarget::VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM) 319 : Fn(Fn), TM(TM), 320 IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {} 321 322 namespace { 323 324 // A slot in a set of virtual tables. The TypeID identifies the set of virtual 325 // tables, and the ByteOffset is the offset in bytes from the address point to 326 // the virtual function pointer. 327 struct VTableSlot { 328 Metadata *TypeID; 329 uint64_t ByteOffset; 330 }; 331 332 } // end anonymous namespace 333 334 namespace llvm { 335 336 template <> struct DenseMapInfo<VTableSlot> { 337 static VTableSlot getEmptyKey() { 338 return {DenseMapInfo<Metadata *>::getEmptyKey(), 339 DenseMapInfo<uint64_t>::getEmptyKey()}; 340 } 341 static VTableSlot getTombstoneKey() { 342 return {DenseMapInfo<Metadata *>::getTombstoneKey(), 343 DenseMapInfo<uint64_t>::getTombstoneKey()}; 344 } 345 static unsigned getHashValue(const VTableSlot &I) { 346 return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^ 347 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 348 } 349 static bool isEqual(const VTableSlot &LHS, 350 const VTableSlot &RHS) { 351 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 352 } 353 }; 354 355 template <> struct DenseMapInfo<VTableSlotSummary> { 356 static VTableSlotSummary getEmptyKey() { 357 return {DenseMapInfo<StringRef>::getEmptyKey(), 358 DenseMapInfo<uint64_t>::getEmptyKey()}; 359 } 360 static VTableSlotSummary getTombstoneKey() { 361 return {DenseMapInfo<StringRef>::getTombstoneKey(), 362 DenseMapInfo<uint64_t>::getTombstoneKey()}; 363 } 364 static unsigned getHashValue(const VTableSlotSummary &I) { 365 return DenseMapInfo<StringRef>::getHashValue(I.TypeID) ^ 366 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 367 } 368 static bool isEqual(const VTableSlotSummary &LHS, 369 const VTableSlotSummary &RHS) { 370 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 371 } 372 }; 373 374 } // end namespace llvm 375 376 namespace { 377 378 // Returns true if the function must be unreachable based on ValueInfo. 379 // 380 // In particular, identifies a function as unreachable in the following 381 // conditions 382 // 1) All summaries are live. 383 // 2) All function summaries indicate it's unreachable 384 bool mustBeUnreachableFunction(ValueInfo TheFnVI) { 385 if ((!TheFnVI) || TheFnVI.getSummaryList().empty()) { 386 // Returns false if ValueInfo is absent, or the summary list is empty 387 // (e.g., function declarations). 388 return false; 389 } 390 391 for (auto &Summary : TheFnVI.getSummaryList()) { 392 // Conservatively returns false if any non-live functions are seen. 393 // In general either all summaries should be live or all should be dead. 394 if (!Summary->isLive()) 395 return false; 396 if (auto *FS = dyn_cast<FunctionSummary>(Summary.get())) { 397 if (!FS->fflags().MustBeUnreachable) 398 return false; 399 } 400 // Do nothing if a non-function has the same GUID (which is rare). 401 // This is correct since non-function summaries are not relevant. 402 } 403 // All function summaries are live and all of them agree that the function is 404 // unreachble. 405 return true; 406 } 407 408 // A virtual call site. VTable is the loaded virtual table pointer, and CS is 409 // the indirect virtual call. 410 struct VirtualCallSite { 411 Value *VTable = nullptr; 412 CallBase &CB; 413 414 // If non-null, this field points to the associated unsafe use count stored in 415 // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description 416 // of that field for details. 417 unsigned *NumUnsafeUses = nullptr; 418 419 void 420 emitRemark(const StringRef OptName, const StringRef TargetName, 421 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) { 422 Function *F = CB.getCaller(); 423 DebugLoc DLoc = CB.getDebugLoc(); 424 BasicBlock *Block = CB.getParent(); 425 426 using namespace ore; 427 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, OptName, DLoc, Block) 428 << NV("Optimization", OptName) 429 << ": devirtualized a call to " 430 << NV("FunctionName", TargetName)); 431 } 432 433 void replaceAndErase( 434 const StringRef OptName, const StringRef TargetName, bool RemarksEnabled, 435 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 436 Value *New) { 437 if (RemarksEnabled) 438 emitRemark(OptName, TargetName, OREGetter); 439 CB.replaceAllUsesWith(New); 440 if (auto *II = dyn_cast<InvokeInst>(&CB)) { 441 BranchInst::Create(II->getNormalDest(), &CB); 442 II->getUnwindDest()->removePredecessor(II->getParent()); 443 } 444 CB.eraseFromParent(); 445 // This use is no longer unsafe. 446 if (NumUnsafeUses) 447 --*NumUnsafeUses; 448 } 449 }; 450 451 // Call site information collected for a specific VTableSlot and possibly a list 452 // of constant integer arguments. The grouping by arguments is handled by the 453 // VTableSlotInfo class. 454 struct CallSiteInfo { 455 /// The set of call sites for this slot. Used during regular LTO and the 456 /// import phase of ThinLTO (as well as the export phase of ThinLTO for any 457 /// call sites that appear in the merged module itself); in each of these 458 /// cases we are directly operating on the call sites at the IR level. 459 std::vector<VirtualCallSite> CallSites; 460 461 /// Whether all call sites represented by this CallSiteInfo, including those 462 /// in summaries, have been devirtualized. This starts off as true because a 463 /// default constructed CallSiteInfo represents no call sites. 464 bool AllCallSitesDevirted = true; 465 466 // These fields are used during the export phase of ThinLTO and reflect 467 // information collected from function summaries. 468 469 /// Whether any function summary contains an llvm.assume(llvm.type.test) for 470 /// this slot. 471 bool SummaryHasTypeTestAssumeUsers = false; 472 473 /// CFI-specific: a vector containing the list of function summaries that use 474 /// the llvm.type.checked.load intrinsic and therefore will require 475 /// resolutions for llvm.type.test in order to implement CFI checks if 476 /// devirtualization was unsuccessful. If devirtualization was successful, the 477 /// pass will clear this vector by calling markDevirt(). If at the end of the 478 /// pass the vector is non-empty, we will need to add a use of llvm.type.test 479 /// to each of the function summaries in the vector. 480 std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers; 481 std::vector<FunctionSummary *> SummaryTypeTestAssumeUsers; 482 483 bool isExported() const { 484 return SummaryHasTypeTestAssumeUsers || 485 !SummaryTypeCheckedLoadUsers.empty(); 486 } 487 488 void addSummaryTypeCheckedLoadUser(FunctionSummary *FS) { 489 SummaryTypeCheckedLoadUsers.push_back(FS); 490 AllCallSitesDevirted = false; 491 } 492 493 void addSummaryTypeTestAssumeUser(FunctionSummary *FS) { 494 SummaryTypeTestAssumeUsers.push_back(FS); 495 SummaryHasTypeTestAssumeUsers = true; 496 AllCallSitesDevirted = false; 497 } 498 499 void markDevirt() { 500 AllCallSitesDevirted = true; 501 502 // As explained in the comment for SummaryTypeCheckedLoadUsers. 503 SummaryTypeCheckedLoadUsers.clear(); 504 } 505 }; 506 507 // Call site information collected for a specific VTableSlot. 508 struct VTableSlotInfo { 509 // The set of call sites which do not have all constant integer arguments 510 // (excluding "this"). 511 CallSiteInfo CSInfo; 512 513 // The set of call sites with all constant integer arguments (excluding 514 // "this"), grouped by argument list. 515 std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo; 516 517 void addCallSite(Value *VTable, CallBase &CB, unsigned *NumUnsafeUses); 518 519 private: 520 CallSiteInfo &findCallSiteInfo(CallBase &CB); 521 }; 522 523 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallBase &CB) { 524 std::vector<uint64_t> Args; 525 auto *CBType = dyn_cast<IntegerType>(CB.getType()); 526 if (!CBType || CBType->getBitWidth() > 64 || CB.arg_empty()) 527 return CSInfo; 528 for (auto &&Arg : drop_begin(CB.args())) { 529 auto *CI = dyn_cast<ConstantInt>(Arg); 530 if (!CI || CI->getBitWidth() > 64) 531 return CSInfo; 532 Args.push_back(CI->getZExtValue()); 533 } 534 return ConstCSInfo[Args]; 535 } 536 537 void VTableSlotInfo::addCallSite(Value *VTable, CallBase &CB, 538 unsigned *NumUnsafeUses) { 539 auto &CSI = findCallSiteInfo(CB); 540 CSI.AllCallSitesDevirted = false; 541 CSI.CallSites.push_back({VTable, CB, NumUnsafeUses}); 542 } 543 544 struct DevirtModule { 545 Module &M; 546 function_ref<AAResults &(Function &)> AARGetter; 547 function_ref<DominatorTree &(Function &)> LookupDomTree; 548 549 ModuleSummaryIndex *ExportSummary; 550 const ModuleSummaryIndex *ImportSummary; 551 552 IntegerType *Int8Ty; 553 PointerType *Int8PtrTy; 554 IntegerType *Int32Ty; 555 IntegerType *Int64Ty; 556 IntegerType *IntPtrTy; 557 /// Sizeless array type, used for imported vtables. This provides a signal 558 /// to analyzers that these imports may alias, as they do for example 559 /// when multiple unique return values occur in the same vtable. 560 ArrayType *Int8Arr0Ty; 561 562 bool RemarksEnabled; 563 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter; 564 565 MapVector<VTableSlot, VTableSlotInfo> CallSlots; 566 567 // Calls that have already been optimized. We may add a call to multiple 568 // VTableSlotInfos if vtable loads are coalesced and need to make sure not to 569 // optimize a call more than once. 570 SmallPtrSet<CallBase *, 8> OptimizedCalls; 571 572 // This map keeps track of the number of "unsafe" uses of a loaded function 573 // pointer. The key is the associated llvm.type.test intrinsic call generated 574 // by this pass. An unsafe use is one that calls the loaded function pointer 575 // directly. Every time we eliminate an unsafe use (for example, by 576 // devirtualizing it or by applying virtual constant propagation), we 577 // decrement the value stored in this map. If a value reaches zero, we can 578 // eliminate the type check by RAUWing the associated llvm.type.test call with 579 // true. 580 std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest; 581 PatternList FunctionsToSkip; 582 583 DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter, 584 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 585 function_ref<DominatorTree &(Function &)> LookupDomTree, 586 ModuleSummaryIndex *ExportSummary, 587 const ModuleSummaryIndex *ImportSummary) 588 : M(M), AARGetter(AARGetter), LookupDomTree(LookupDomTree), 589 ExportSummary(ExportSummary), ImportSummary(ImportSummary), 590 Int8Ty(Type::getInt8Ty(M.getContext())), 591 Int8PtrTy(Type::getInt8PtrTy(M.getContext())), 592 Int32Ty(Type::getInt32Ty(M.getContext())), 593 Int64Ty(Type::getInt64Ty(M.getContext())), 594 IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)), 595 Int8Arr0Ty(ArrayType::get(Type::getInt8Ty(M.getContext()), 0)), 596 RemarksEnabled(areRemarksEnabled()), OREGetter(OREGetter) { 597 assert(!(ExportSummary && ImportSummary)); 598 FunctionsToSkip.init(SkipFunctionNames); 599 } 600 601 bool areRemarksEnabled(); 602 603 void 604 scanTypeTestUsers(Function *TypeTestFunc, 605 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap); 606 void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc); 607 608 void buildTypeIdentifierMap( 609 std::vector<VTableBits> &Bits, 610 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap); 611 612 bool 613 tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot, 614 const std::set<TypeMemberInfo> &TypeMemberInfos, 615 uint64_t ByteOffset, 616 ModuleSummaryIndex *ExportSummary); 617 618 void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn, 619 bool &IsExported); 620 bool trySingleImplDevirt(ModuleSummaryIndex *ExportSummary, 621 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 622 VTableSlotInfo &SlotInfo, 623 WholeProgramDevirtResolution *Res); 624 625 void applyICallBranchFunnel(VTableSlotInfo &SlotInfo, Constant *JT, 626 bool &IsExported); 627 void tryICallBranchFunnel(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 628 VTableSlotInfo &SlotInfo, 629 WholeProgramDevirtResolution *Res, VTableSlot Slot); 630 631 bool tryEvaluateFunctionsWithArgs( 632 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 633 ArrayRef<uint64_t> Args); 634 635 void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 636 uint64_t TheRetVal); 637 bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 638 CallSiteInfo &CSInfo, 639 WholeProgramDevirtResolution::ByArg *Res); 640 641 // Returns the global symbol name that is used to export information about the 642 // given vtable slot and list of arguments. 643 std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args, 644 StringRef Name); 645 646 bool shouldExportConstantsAsAbsoluteSymbols(); 647 648 // This function is called during the export phase to create a symbol 649 // definition containing information about the given vtable slot and list of 650 // arguments. 651 void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 652 Constant *C); 653 void exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 654 uint32_t Const, uint32_t &Storage); 655 656 // This function is called during the import phase to create a reference to 657 // the symbol definition created during the export phase. 658 Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 659 StringRef Name); 660 Constant *importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 661 StringRef Name, IntegerType *IntTy, 662 uint32_t Storage); 663 664 Constant *getMemberAddr(const TypeMemberInfo *M); 665 666 void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne, 667 Constant *UniqueMemberAddr); 668 bool tryUniqueRetValOpt(unsigned BitWidth, 669 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 670 CallSiteInfo &CSInfo, 671 WholeProgramDevirtResolution::ByArg *Res, 672 VTableSlot Slot, ArrayRef<uint64_t> Args); 673 674 void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 675 Constant *Byte, Constant *Bit); 676 bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 677 VTableSlotInfo &SlotInfo, 678 WholeProgramDevirtResolution *Res, VTableSlot Slot); 679 680 void rebuildGlobal(VTableBits &B); 681 682 // Apply the summary resolution for Slot to all virtual calls in SlotInfo. 683 void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo); 684 685 // If we were able to eliminate all unsafe uses for a type checked load, 686 // eliminate the associated type tests by replacing them with true. 687 void removeRedundantTypeTests(); 688 689 bool run(); 690 691 // Look up the corresponding ValueInfo entry of `TheFn` in `ExportSummary`. 692 // 693 // Caller guarantees that `ExportSummary` is not nullptr. 694 static ValueInfo lookUpFunctionValueInfo(Function *TheFn, 695 ModuleSummaryIndex *ExportSummary); 696 697 // Returns true if the function definition must be unreachable. 698 // 699 // Note if this helper function returns true, `F` is guaranteed 700 // to be unreachable; if it returns false, `F` might still 701 // be unreachable but not covered by this helper function. 702 // 703 // Implementation-wise, if function definition is present, IR is analyzed; if 704 // not, look up function flags from ExportSummary as a fallback. 705 static bool mustBeUnreachableFunction(Function *const F, 706 ModuleSummaryIndex *ExportSummary); 707 708 // Lower the module using the action and summary passed as command line 709 // arguments. For testing purposes only. 710 static bool 711 runForTesting(Module &M, function_ref<AAResults &(Function &)> AARGetter, 712 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 713 function_ref<DominatorTree &(Function &)> LookupDomTree); 714 }; 715 716 struct DevirtIndex { 717 ModuleSummaryIndex &ExportSummary; 718 // The set in which to record GUIDs exported from their module by 719 // devirtualization, used by client to ensure they are not internalized. 720 std::set<GlobalValue::GUID> &ExportedGUIDs; 721 // A map in which to record the information necessary to locate the WPD 722 // resolution for local targets in case they are exported by cross module 723 // importing. 724 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap; 725 726 MapVector<VTableSlotSummary, VTableSlotInfo> CallSlots; 727 728 PatternList FunctionsToSkip; 729 730 DevirtIndex( 731 ModuleSummaryIndex &ExportSummary, 732 std::set<GlobalValue::GUID> &ExportedGUIDs, 733 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) 734 : ExportSummary(ExportSummary), ExportedGUIDs(ExportedGUIDs), 735 LocalWPDTargetsMap(LocalWPDTargetsMap) { 736 FunctionsToSkip.init(SkipFunctionNames); 737 } 738 739 bool tryFindVirtualCallTargets(std::vector<ValueInfo> &TargetsForSlot, 740 const TypeIdCompatibleVtableInfo TIdInfo, 741 uint64_t ByteOffset); 742 743 bool trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 744 VTableSlotSummary &SlotSummary, 745 VTableSlotInfo &SlotInfo, 746 WholeProgramDevirtResolution *Res, 747 std::set<ValueInfo> &DevirtTargets); 748 749 void run(); 750 }; 751 752 struct WholeProgramDevirt : public ModulePass { 753 static char ID; 754 755 bool UseCommandLine = false; 756 757 ModuleSummaryIndex *ExportSummary = nullptr; 758 const ModuleSummaryIndex *ImportSummary = nullptr; 759 760 WholeProgramDevirt() : ModulePass(ID), UseCommandLine(true) { 761 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry()); 762 } 763 764 WholeProgramDevirt(ModuleSummaryIndex *ExportSummary, 765 const ModuleSummaryIndex *ImportSummary) 766 : ModulePass(ID), ExportSummary(ExportSummary), 767 ImportSummary(ImportSummary) { 768 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry()); 769 } 770 771 bool runOnModule(Module &M) override { 772 if (skipModule(M)) 773 return false; 774 775 // In the new pass manager, we can request the optimization 776 // remark emitter pass on a per-function-basis, which the 777 // OREGetter will do for us. 778 // In the old pass manager, this is harder, so we just build 779 // an optimization remark emitter on the fly, when we need it. 780 std::unique_ptr<OptimizationRemarkEmitter> ORE; 781 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { 782 ORE = std::make_unique<OptimizationRemarkEmitter>(F); 783 return *ORE; 784 }; 785 786 auto LookupDomTree = [this](Function &F) -> DominatorTree & { 787 return this->getAnalysis<DominatorTreeWrapperPass>(F).getDomTree(); 788 }; 789 790 if (UseCommandLine) 791 return DevirtModule::runForTesting(M, LegacyAARGetter(*this), OREGetter, 792 LookupDomTree); 793 794 return DevirtModule(M, LegacyAARGetter(*this), OREGetter, LookupDomTree, 795 ExportSummary, ImportSummary) 796 .run(); 797 } 798 799 void getAnalysisUsage(AnalysisUsage &AU) const override { 800 AU.addRequired<AssumptionCacheTracker>(); 801 AU.addRequired<TargetLibraryInfoWrapperPass>(); 802 AU.addRequired<DominatorTreeWrapperPass>(); 803 } 804 }; 805 806 } // end anonymous namespace 807 808 INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt", 809 "Whole program devirtualization", false, false) 810 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 811 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 812 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 813 INITIALIZE_PASS_END(WholeProgramDevirt, "wholeprogramdevirt", 814 "Whole program devirtualization", false, false) 815 char WholeProgramDevirt::ID = 0; 816 817 ModulePass * 818 llvm::createWholeProgramDevirtPass(ModuleSummaryIndex *ExportSummary, 819 const ModuleSummaryIndex *ImportSummary) { 820 return new WholeProgramDevirt(ExportSummary, ImportSummary); 821 } 822 823 PreservedAnalyses WholeProgramDevirtPass::run(Module &M, 824 ModuleAnalysisManager &AM) { 825 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 826 auto AARGetter = [&](Function &F) -> AAResults & { 827 return FAM.getResult<AAManager>(F); 828 }; 829 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { 830 return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F); 831 }; 832 auto LookupDomTree = [&FAM](Function &F) -> DominatorTree & { 833 return FAM.getResult<DominatorTreeAnalysis>(F); 834 }; 835 if (UseCommandLine) { 836 if (DevirtModule::runForTesting(M, AARGetter, OREGetter, LookupDomTree)) 837 return PreservedAnalyses::all(); 838 return PreservedAnalyses::none(); 839 } 840 if (!DevirtModule(M, AARGetter, OREGetter, LookupDomTree, ExportSummary, 841 ImportSummary) 842 .run()) 843 return PreservedAnalyses::all(); 844 return PreservedAnalyses::none(); 845 } 846 847 // Enable whole program visibility if enabled by client (e.g. linker) or 848 // internal option, and not force disabled. 849 static bool hasWholeProgramVisibility(bool WholeProgramVisibilityEnabledInLTO) { 850 return (WholeProgramVisibilityEnabledInLTO || WholeProgramVisibility) && 851 !DisableWholeProgramVisibility; 852 } 853 854 namespace llvm { 855 856 /// If whole program visibility asserted, then upgrade all public vcall 857 /// visibility metadata on vtable definitions to linkage unit visibility in 858 /// Module IR (for regular or hybrid LTO). 859 void updateVCallVisibilityInModule( 860 Module &M, bool WholeProgramVisibilityEnabledInLTO, 861 const DenseSet<GlobalValue::GUID> &DynamicExportSymbols) { 862 if (!hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO)) 863 return; 864 for (GlobalVariable &GV : M.globals()) 865 // Add linkage unit visibility to any variable with type metadata, which are 866 // the vtable definitions. We won't have an existing vcall_visibility 867 // metadata on vtable definitions with public visibility. 868 if (GV.hasMetadata(LLVMContext::MD_type) && 869 GV.getVCallVisibility() == GlobalObject::VCallVisibilityPublic && 870 // Don't upgrade the visibility for symbols exported to the dynamic 871 // linker, as we have no information on their eventual use. 872 !DynamicExportSymbols.count(GV.getGUID())) 873 GV.setVCallVisibilityMetadata(GlobalObject::VCallVisibilityLinkageUnit); 874 } 875 876 /// If whole program visibility asserted, then upgrade all public vcall 877 /// visibility metadata on vtable definition summaries to linkage unit 878 /// visibility in Module summary index (for ThinLTO). 879 void updateVCallVisibilityInIndex( 880 ModuleSummaryIndex &Index, bool WholeProgramVisibilityEnabledInLTO, 881 const DenseSet<GlobalValue::GUID> &DynamicExportSymbols) { 882 if (!hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO)) 883 return; 884 for (auto &P : Index) { 885 // Don't upgrade the visibility for symbols exported to the dynamic 886 // linker, as we have no information on their eventual use. 887 if (DynamicExportSymbols.count(P.first)) 888 continue; 889 for (auto &S : P.second.SummaryList) { 890 auto *GVar = dyn_cast<GlobalVarSummary>(S.get()); 891 if (!GVar || 892 GVar->getVCallVisibility() != GlobalObject::VCallVisibilityPublic) 893 continue; 894 GVar->setVCallVisibility(GlobalObject::VCallVisibilityLinkageUnit); 895 } 896 } 897 } 898 899 void runWholeProgramDevirtOnIndex( 900 ModuleSummaryIndex &Summary, std::set<GlobalValue::GUID> &ExportedGUIDs, 901 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 902 DevirtIndex(Summary, ExportedGUIDs, LocalWPDTargetsMap).run(); 903 } 904 905 void updateIndexWPDForExports( 906 ModuleSummaryIndex &Summary, 907 function_ref<bool(StringRef, ValueInfo)> isExported, 908 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 909 for (auto &T : LocalWPDTargetsMap) { 910 auto &VI = T.first; 911 // This was enforced earlier during trySingleImplDevirt. 912 assert(VI.getSummaryList().size() == 1 && 913 "Devirt of local target has more than one copy"); 914 auto &S = VI.getSummaryList()[0]; 915 if (!isExported(S->modulePath(), VI)) 916 continue; 917 918 // It's been exported by a cross module import. 919 for (auto &SlotSummary : T.second) { 920 auto *TIdSum = Summary.getTypeIdSummary(SlotSummary.TypeID); 921 assert(TIdSum); 922 auto WPDRes = TIdSum->WPDRes.find(SlotSummary.ByteOffset); 923 assert(WPDRes != TIdSum->WPDRes.end()); 924 WPDRes->second.SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 925 WPDRes->second.SingleImplName, 926 Summary.getModuleHash(S->modulePath())); 927 } 928 } 929 } 930 931 } // end namespace llvm 932 933 static Error checkCombinedSummaryForTesting(ModuleSummaryIndex *Summary) { 934 // Check that summary index contains regular LTO module when performing 935 // export to prevent occasional use of index from pure ThinLTO compilation 936 // (-fno-split-lto-module). This kind of summary index is passed to 937 // DevirtIndex::run, not to DevirtModule::run used by opt/runForTesting. 938 const auto &ModPaths = Summary->modulePaths(); 939 if (ClSummaryAction != PassSummaryAction::Import && 940 ModPaths.find(ModuleSummaryIndex::getRegularLTOModuleName()) == 941 ModPaths.end()) 942 return createStringError( 943 errc::invalid_argument, 944 "combined summary should contain Regular LTO module"); 945 return ErrorSuccess(); 946 } 947 948 bool DevirtModule::runForTesting( 949 Module &M, function_ref<AAResults &(Function &)> AARGetter, 950 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 951 function_ref<DominatorTree &(Function &)> LookupDomTree) { 952 std::unique_ptr<ModuleSummaryIndex> Summary = 953 std::make_unique<ModuleSummaryIndex>(/*HaveGVs=*/false); 954 955 // Handle the command-line summary arguments. This code is for testing 956 // purposes only, so we handle errors directly. 957 if (!ClReadSummary.empty()) { 958 ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary + 959 ": "); 960 auto ReadSummaryFile = 961 ExitOnErr(errorOrToExpected(MemoryBuffer::getFile(ClReadSummary))); 962 if (Expected<std::unique_ptr<ModuleSummaryIndex>> SummaryOrErr = 963 getModuleSummaryIndex(*ReadSummaryFile)) { 964 Summary = std::move(*SummaryOrErr); 965 ExitOnErr(checkCombinedSummaryForTesting(Summary.get())); 966 } else { 967 // Try YAML if we've failed with bitcode. 968 consumeError(SummaryOrErr.takeError()); 969 yaml::Input In(ReadSummaryFile->getBuffer()); 970 In >> *Summary; 971 ExitOnErr(errorCodeToError(In.error())); 972 } 973 } 974 975 bool Changed = 976 DevirtModule(M, AARGetter, OREGetter, LookupDomTree, 977 ClSummaryAction == PassSummaryAction::Export ? Summary.get() 978 : nullptr, 979 ClSummaryAction == PassSummaryAction::Import ? Summary.get() 980 : nullptr) 981 .run(); 982 983 if (!ClWriteSummary.empty()) { 984 ExitOnError ExitOnErr( 985 "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": "); 986 std::error_code EC; 987 if (StringRef(ClWriteSummary).endswith(".bc")) { 988 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_None); 989 ExitOnErr(errorCodeToError(EC)); 990 writeIndexToFile(*Summary, OS); 991 } else { 992 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_TextWithCRLF); 993 ExitOnErr(errorCodeToError(EC)); 994 yaml::Output Out(OS); 995 Out << *Summary; 996 } 997 } 998 999 return Changed; 1000 } 1001 1002 void DevirtModule::buildTypeIdentifierMap( 1003 std::vector<VTableBits> &Bits, 1004 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 1005 DenseMap<GlobalVariable *, VTableBits *> GVToBits; 1006 Bits.reserve(M.getGlobalList().size()); 1007 SmallVector<MDNode *, 2> Types; 1008 for (GlobalVariable &GV : M.globals()) { 1009 Types.clear(); 1010 GV.getMetadata(LLVMContext::MD_type, Types); 1011 if (GV.isDeclaration() || Types.empty()) 1012 continue; 1013 1014 VTableBits *&BitsPtr = GVToBits[&GV]; 1015 if (!BitsPtr) { 1016 Bits.emplace_back(); 1017 Bits.back().GV = &GV; 1018 Bits.back().ObjectSize = 1019 M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType()); 1020 BitsPtr = &Bits.back(); 1021 } 1022 1023 for (MDNode *Type : Types) { 1024 auto TypeID = Type->getOperand(1).get(); 1025 1026 uint64_t Offset = 1027 cast<ConstantInt>( 1028 cast<ConstantAsMetadata>(Type->getOperand(0))->getValue()) 1029 ->getZExtValue(); 1030 1031 TypeIdMap[TypeID].insert({BitsPtr, Offset}); 1032 } 1033 } 1034 } 1035 1036 bool DevirtModule::tryFindVirtualCallTargets( 1037 std::vector<VirtualCallTarget> &TargetsForSlot, 1038 const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset, 1039 ModuleSummaryIndex *ExportSummary) { 1040 for (const TypeMemberInfo &TM : TypeMemberInfos) { 1041 if (!TM.Bits->GV->isConstant()) 1042 return false; 1043 1044 // We cannot perform whole program devirtualization analysis on a vtable 1045 // with public LTO visibility. 1046 if (TM.Bits->GV->getVCallVisibility() == 1047 GlobalObject::VCallVisibilityPublic) 1048 return false; 1049 1050 Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(), 1051 TM.Offset + ByteOffset, M); 1052 if (!Ptr) 1053 return false; 1054 1055 auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts()); 1056 if (!Fn) 1057 return false; 1058 1059 if (FunctionsToSkip.match(Fn->getName())) 1060 return false; 1061 1062 // We can disregard __cxa_pure_virtual as a possible call target, as 1063 // calls to pure virtuals are UB. 1064 if (Fn->getName() == "__cxa_pure_virtual") 1065 continue; 1066 1067 // We can disregard unreachable functions as possible call targets, as 1068 // unreachable functions shouldn't be called. 1069 if (mustBeUnreachableFunction(Fn, ExportSummary)) 1070 continue; 1071 1072 TargetsForSlot.push_back({Fn, &TM}); 1073 } 1074 1075 // Give up if we couldn't find any targets. 1076 return !TargetsForSlot.empty(); 1077 } 1078 1079 bool DevirtIndex::tryFindVirtualCallTargets( 1080 std::vector<ValueInfo> &TargetsForSlot, const TypeIdCompatibleVtableInfo TIdInfo, 1081 uint64_t ByteOffset) { 1082 for (const TypeIdOffsetVtableInfo &P : TIdInfo) { 1083 // Find a representative copy of the vtable initializer. 1084 // We can have multiple available_externally, linkonce_odr and weak_odr 1085 // vtable initializers. We can also have multiple external vtable 1086 // initializers in the case of comdats, which we cannot check here. 1087 // The linker should give an error in this case. 1088 // 1089 // Also, handle the case of same-named local Vtables with the same path 1090 // and therefore the same GUID. This can happen if there isn't enough 1091 // distinguishing path when compiling the source file. In that case we 1092 // conservatively return false early. 1093 const GlobalVarSummary *VS = nullptr; 1094 bool LocalFound = false; 1095 for (auto &S : P.VTableVI.getSummaryList()) { 1096 if (GlobalValue::isLocalLinkage(S->linkage())) { 1097 if (LocalFound) 1098 return false; 1099 LocalFound = true; 1100 } 1101 auto *CurVS = cast<GlobalVarSummary>(S->getBaseObject()); 1102 if (!CurVS->vTableFuncs().empty() || 1103 // Previously clang did not attach the necessary type metadata to 1104 // available_externally vtables, in which case there would not 1105 // be any vtable functions listed in the summary and we need 1106 // to treat this case conservatively (in case the bitcode is old). 1107 // However, we will also not have any vtable functions in the 1108 // case of a pure virtual base class. In that case we do want 1109 // to set VS to avoid treating it conservatively. 1110 !GlobalValue::isAvailableExternallyLinkage(S->linkage())) { 1111 VS = CurVS; 1112 // We cannot perform whole program devirtualization analysis on a vtable 1113 // with public LTO visibility. 1114 if (VS->getVCallVisibility() == GlobalObject::VCallVisibilityPublic) 1115 return false; 1116 } 1117 } 1118 // There will be no VS if all copies are available_externally having no 1119 // type metadata. In that case we can't safely perform WPD. 1120 if (!VS) 1121 return false; 1122 if (!VS->isLive()) 1123 continue; 1124 for (auto VTP : VS->vTableFuncs()) { 1125 if (VTP.VTableOffset != P.AddressPointOffset + ByteOffset) 1126 continue; 1127 1128 if (mustBeUnreachableFunction(VTP.FuncVI)) 1129 continue; 1130 1131 TargetsForSlot.push_back(VTP.FuncVI); 1132 } 1133 } 1134 1135 // Give up if we couldn't find any targets. 1136 return !TargetsForSlot.empty(); 1137 } 1138 1139 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo, 1140 Constant *TheFn, bool &IsExported) { 1141 // Don't devirtualize function if we're told to skip it 1142 // in -wholeprogramdevirt-skip. 1143 if (FunctionsToSkip.match(TheFn->stripPointerCasts()->getName())) 1144 return; 1145 auto Apply = [&](CallSiteInfo &CSInfo) { 1146 for (auto &&VCallSite : CSInfo.CallSites) { 1147 if (!OptimizedCalls.insert(&VCallSite.CB).second) 1148 continue; 1149 1150 if (RemarksEnabled) 1151 VCallSite.emitRemark("single-impl", 1152 TheFn->stripPointerCasts()->getName(), OREGetter); 1153 NumSingleImpl++; 1154 auto &CB = VCallSite.CB; 1155 assert(!CB.getCalledFunction() && "devirtualizing direct call?"); 1156 IRBuilder<> Builder(&CB); 1157 Value *Callee = 1158 Builder.CreateBitCast(TheFn, CB.getCalledOperand()->getType()); 1159 1160 // If trap checking is enabled, add support to compare the virtual 1161 // function pointer to the devirtualized target. In case of a mismatch, 1162 // perform a debug trap. 1163 if (DevirtCheckMode == WPDCheckMode::Trap) { 1164 auto *Cond = Builder.CreateICmpNE(CB.getCalledOperand(), Callee); 1165 Instruction *ThenTerm = 1166 SplitBlockAndInsertIfThen(Cond, &CB, /*Unreachable=*/false); 1167 Builder.SetInsertPoint(ThenTerm); 1168 Function *TrapFn = Intrinsic::getDeclaration(&M, Intrinsic::debugtrap); 1169 auto *CallTrap = Builder.CreateCall(TrapFn); 1170 CallTrap->setDebugLoc(CB.getDebugLoc()); 1171 } 1172 1173 // If fallback checking is enabled, add support to compare the virtual 1174 // function pointer to the devirtualized target. In case of a mismatch, 1175 // fall back to indirect call. 1176 if (DevirtCheckMode == WPDCheckMode::Fallback) { 1177 MDNode *Weights = 1178 MDBuilder(M.getContext()).createBranchWeights((1U << 20) - 1, 1); 1179 // Version the indirect call site. If the called value is equal to the 1180 // given callee, 'NewInst' will be executed, otherwise the original call 1181 // site will be executed. 1182 CallBase &NewInst = versionCallSite(CB, Callee, Weights); 1183 NewInst.setCalledOperand(Callee); 1184 // Since the new call site is direct, we must clear metadata that 1185 // is only appropriate for indirect calls. This includes !prof and 1186 // !callees metadata. 1187 NewInst.setMetadata(LLVMContext::MD_prof, nullptr); 1188 NewInst.setMetadata(LLVMContext::MD_callees, nullptr); 1189 // Additionally, we should remove them from the fallback indirect call, 1190 // so that we don't attempt to perform indirect call promotion later. 1191 CB.setMetadata(LLVMContext::MD_prof, nullptr); 1192 CB.setMetadata(LLVMContext::MD_callees, nullptr); 1193 } 1194 1195 // In either trapping or non-checking mode, devirtualize original call. 1196 else { 1197 // Devirtualize unconditionally. 1198 CB.setCalledOperand(Callee); 1199 // Since the call site is now direct, we must clear metadata that 1200 // is only appropriate for indirect calls. This includes !prof and 1201 // !callees metadata. 1202 CB.setMetadata(LLVMContext::MD_prof, nullptr); 1203 CB.setMetadata(LLVMContext::MD_callees, nullptr); 1204 } 1205 1206 // This use is no longer unsafe. 1207 if (VCallSite.NumUnsafeUses) 1208 --*VCallSite.NumUnsafeUses; 1209 } 1210 if (CSInfo.isExported()) 1211 IsExported = true; 1212 CSInfo.markDevirt(); 1213 }; 1214 Apply(SlotInfo.CSInfo); 1215 for (auto &P : SlotInfo.ConstCSInfo) 1216 Apply(P.second); 1217 } 1218 1219 static bool AddCalls(VTableSlotInfo &SlotInfo, const ValueInfo &Callee) { 1220 // We can't add calls if we haven't seen a definition 1221 if (Callee.getSummaryList().empty()) 1222 return false; 1223 1224 // Insert calls into the summary index so that the devirtualized targets 1225 // are eligible for import. 1226 // FIXME: Annotate type tests with hotness. For now, mark these as hot 1227 // to better ensure we have the opportunity to inline them. 1228 bool IsExported = false; 1229 auto &S = Callee.getSummaryList()[0]; 1230 CalleeInfo CI(CalleeInfo::HotnessType::Hot, /* RelBF = */ 0); 1231 auto AddCalls = [&](CallSiteInfo &CSInfo) { 1232 for (auto *FS : CSInfo.SummaryTypeCheckedLoadUsers) { 1233 FS->addCall({Callee, CI}); 1234 IsExported |= S->modulePath() != FS->modulePath(); 1235 } 1236 for (auto *FS : CSInfo.SummaryTypeTestAssumeUsers) { 1237 FS->addCall({Callee, CI}); 1238 IsExported |= S->modulePath() != FS->modulePath(); 1239 } 1240 }; 1241 AddCalls(SlotInfo.CSInfo); 1242 for (auto &P : SlotInfo.ConstCSInfo) 1243 AddCalls(P.second); 1244 return IsExported; 1245 } 1246 1247 bool DevirtModule::trySingleImplDevirt( 1248 ModuleSummaryIndex *ExportSummary, 1249 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1250 WholeProgramDevirtResolution *Res) { 1251 // See if the program contains a single implementation of this virtual 1252 // function. 1253 Function *TheFn = TargetsForSlot[0].Fn; 1254 for (auto &&Target : TargetsForSlot) 1255 if (TheFn != Target.Fn) 1256 return false; 1257 1258 // If so, update each call site to call that implementation directly. 1259 if (RemarksEnabled || AreStatisticsEnabled()) 1260 TargetsForSlot[0].WasDevirt = true; 1261 1262 bool IsExported = false; 1263 applySingleImplDevirt(SlotInfo, TheFn, IsExported); 1264 if (!IsExported) 1265 return false; 1266 1267 // If the only implementation has local linkage, we must promote to external 1268 // to make it visible to thin LTO objects. We can only get here during the 1269 // ThinLTO export phase. 1270 if (TheFn->hasLocalLinkage()) { 1271 std::string NewName = (TheFn->getName() + ".llvm.merged").str(); 1272 1273 // Since we are renaming the function, any comdats with the same name must 1274 // also be renamed. This is required when targeting COFF, as the comdat name 1275 // must match one of the names of the symbols in the comdat. 1276 if (Comdat *C = TheFn->getComdat()) { 1277 if (C->getName() == TheFn->getName()) { 1278 Comdat *NewC = M.getOrInsertComdat(NewName); 1279 NewC->setSelectionKind(C->getSelectionKind()); 1280 for (GlobalObject &GO : M.global_objects()) 1281 if (GO.getComdat() == C) 1282 GO.setComdat(NewC); 1283 } 1284 } 1285 1286 TheFn->setLinkage(GlobalValue::ExternalLinkage); 1287 TheFn->setVisibility(GlobalValue::HiddenVisibility); 1288 TheFn->setName(NewName); 1289 } 1290 if (ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFn->getGUID())) 1291 // Any needed promotion of 'TheFn' has already been done during 1292 // LTO unit split, so we can ignore return value of AddCalls. 1293 AddCalls(SlotInfo, TheFnVI); 1294 1295 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1296 Res->SingleImplName = std::string(TheFn->getName()); 1297 1298 return true; 1299 } 1300 1301 bool DevirtIndex::trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 1302 VTableSlotSummary &SlotSummary, 1303 VTableSlotInfo &SlotInfo, 1304 WholeProgramDevirtResolution *Res, 1305 std::set<ValueInfo> &DevirtTargets) { 1306 // See if the program contains a single implementation of this virtual 1307 // function. 1308 auto TheFn = TargetsForSlot[0]; 1309 for (auto &&Target : TargetsForSlot) 1310 if (TheFn != Target) 1311 return false; 1312 1313 // Don't devirtualize if we don't have target definition. 1314 auto Size = TheFn.getSummaryList().size(); 1315 if (!Size) 1316 return false; 1317 1318 // Don't devirtualize function if we're told to skip it 1319 // in -wholeprogramdevirt-skip. 1320 if (FunctionsToSkip.match(TheFn.name())) 1321 return false; 1322 1323 // If the summary list contains multiple summaries where at least one is 1324 // a local, give up, as we won't know which (possibly promoted) name to use. 1325 for (auto &S : TheFn.getSummaryList()) 1326 if (GlobalValue::isLocalLinkage(S->linkage()) && Size > 1) 1327 return false; 1328 1329 // Collect functions devirtualized at least for one call site for stats. 1330 if (PrintSummaryDevirt || AreStatisticsEnabled()) 1331 DevirtTargets.insert(TheFn); 1332 1333 auto &S = TheFn.getSummaryList()[0]; 1334 bool IsExported = AddCalls(SlotInfo, TheFn); 1335 if (IsExported) 1336 ExportedGUIDs.insert(TheFn.getGUID()); 1337 1338 // Record in summary for use in devirtualization during the ThinLTO import 1339 // step. 1340 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1341 if (GlobalValue::isLocalLinkage(S->linkage())) { 1342 if (IsExported) 1343 // If target is a local function and we are exporting it by 1344 // devirtualizing a call in another module, we need to record the 1345 // promoted name. 1346 Res->SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 1347 TheFn.name(), ExportSummary.getModuleHash(S->modulePath())); 1348 else { 1349 LocalWPDTargetsMap[TheFn].push_back(SlotSummary); 1350 Res->SingleImplName = std::string(TheFn.name()); 1351 } 1352 } else 1353 Res->SingleImplName = std::string(TheFn.name()); 1354 1355 // Name will be empty if this thin link driven off of serialized combined 1356 // index (e.g. llvm-lto). However, WPD is not supported/invoked for the 1357 // legacy LTO API anyway. 1358 assert(!Res->SingleImplName.empty()); 1359 1360 return true; 1361 } 1362 1363 void DevirtModule::tryICallBranchFunnel( 1364 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1365 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1366 Triple T(M.getTargetTriple()); 1367 if (T.getArch() != Triple::x86_64) 1368 return; 1369 1370 if (TargetsForSlot.size() > ClThreshold) 1371 return; 1372 1373 bool HasNonDevirt = !SlotInfo.CSInfo.AllCallSitesDevirted; 1374 if (!HasNonDevirt) 1375 for (auto &P : SlotInfo.ConstCSInfo) 1376 if (!P.second.AllCallSitesDevirted) { 1377 HasNonDevirt = true; 1378 break; 1379 } 1380 1381 if (!HasNonDevirt) 1382 return; 1383 1384 FunctionType *FT = 1385 FunctionType::get(Type::getVoidTy(M.getContext()), {Int8PtrTy}, true); 1386 Function *JT; 1387 if (isa<MDString>(Slot.TypeID)) { 1388 JT = Function::Create(FT, Function::ExternalLinkage, 1389 M.getDataLayout().getProgramAddressSpace(), 1390 getGlobalName(Slot, {}, "branch_funnel"), &M); 1391 JT->setVisibility(GlobalValue::HiddenVisibility); 1392 } else { 1393 JT = Function::Create(FT, Function::InternalLinkage, 1394 M.getDataLayout().getProgramAddressSpace(), 1395 "branch_funnel", &M); 1396 } 1397 JT->addParamAttr(0, Attribute::Nest); 1398 1399 std::vector<Value *> JTArgs; 1400 JTArgs.push_back(JT->arg_begin()); 1401 for (auto &T : TargetsForSlot) { 1402 JTArgs.push_back(getMemberAddr(T.TM)); 1403 JTArgs.push_back(T.Fn); 1404 } 1405 1406 BasicBlock *BB = BasicBlock::Create(M.getContext(), "", JT, nullptr); 1407 Function *Intr = 1408 Intrinsic::getDeclaration(&M, llvm::Intrinsic::icall_branch_funnel, {}); 1409 1410 auto *CI = CallInst::Create(Intr, JTArgs, "", BB); 1411 CI->setTailCallKind(CallInst::TCK_MustTail); 1412 ReturnInst::Create(M.getContext(), nullptr, BB); 1413 1414 bool IsExported = false; 1415 applyICallBranchFunnel(SlotInfo, JT, IsExported); 1416 if (IsExported) 1417 Res->TheKind = WholeProgramDevirtResolution::BranchFunnel; 1418 } 1419 1420 void DevirtModule::applyICallBranchFunnel(VTableSlotInfo &SlotInfo, 1421 Constant *JT, bool &IsExported) { 1422 auto Apply = [&](CallSiteInfo &CSInfo) { 1423 if (CSInfo.isExported()) 1424 IsExported = true; 1425 if (CSInfo.AllCallSitesDevirted) 1426 return; 1427 for (auto &&VCallSite : CSInfo.CallSites) { 1428 CallBase &CB = VCallSite.CB; 1429 1430 // Jump tables are only profitable if the retpoline mitigation is enabled. 1431 Attribute FSAttr = CB.getCaller()->getFnAttribute("target-features"); 1432 if (!FSAttr.isValid() || 1433 !FSAttr.getValueAsString().contains("+retpoline")) 1434 continue; 1435 1436 NumBranchFunnel++; 1437 if (RemarksEnabled) 1438 VCallSite.emitRemark("branch-funnel", 1439 JT->stripPointerCasts()->getName(), OREGetter); 1440 1441 // Pass the address of the vtable in the nest register, which is r10 on 1442 // x86_64. 1443 std::vector<Type *> NewArgs; 1444 NewArgs.push_back(Int8PtrTy); 1445 append_range(NewArgs, CB.getFunctionType()->params()); 1446 FunctionType *NewFT = 1447 FunctionType::get(CB.getFunctionType()->getReturnType(), NewArgs, 1448 CB.getFunctionType()->isVarArg()); 1449 PointerType *NewFTPtr = PointerType::getUnqual(NewFT); 1450 1451 IRBuilder<> IRB(&CB); 1452 std::vector<Value *> Args; 1453 Args.push_back(IRB.CreateBitCast(VCallSite.VTable, Int8PtrTy)); 1454 llvm::append_range(Args, CB.args()); 1455 1456 CallBase *NewCS = nullptr; 1457 if (isa<CallInst>(CB)) 1458 NewCS = IRB.CreateCall(NewFT, IRB.CreateBitCast(JT, NewFTPtr), Args); 1459 else 1460 NewCS = IRB.CreateInvoke(NewFT, IRB.CreateBitCast(JT, NewFTPtr), 1461 cast<InvokeInst>(CB).getNormalDest(), 1462 cast<InvokeInst>(CB).getUnwindDest(), Args); 1463 NewCS->setCallingConv(CB.getCallingConv()); 1464 1465 AttributeList Attrs = CB.getAttributes(); 1466 std::vector<AttributeSet> NewArgAttrs; 1467 NewArgAttrs.push_back(AttributeSet::get( 1468 M.getContext(), ArrayRef<Attribute>{Attribute::get( 1469 M.getContext(), Attribute::Nest)})); 1470 for (unsigned I = 0; I + 2 < Attrs.getNumAttrSets(); ++I) 1471 NewArgAttrs.push_back(Attrs.getParamAttrs(I)); 1472 NewCS->setAttributes( 1473 AttributeList::get(M.getContext(), Attrs.getFnAttrs(), 1474 Attrs.getRetAttrs(), NewArgAttrs)); 1475 1476 CB.replaceAllUsesWith(NewCS); 1477 CB.eraseFromParent(); 1478 1479 // This use is no longer unsafe. 1480 if (VCallSite.NumUnsafeUses) 1481 --*VCallSite.NumUnsafeUses; 1482 } 1483 // Don't mark as devirtualized because there may be callers compiled without 1484 // retpoline mitigation, which would mean that they are lowered to 1485 // llvm.type.test and therefore require an llvm.type.test resolution for the 1486 // type identifier. 1487 }; 1488 Apply(SlotInfo.CSInfo); 1489 for (auto &P : SlotInfo.ConstCSInfo) 1490 Apply(P.second); 1491 } 1492 1493 bool DevirtModule::tryEvaluateFunctionsWithArgs( 1494 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1495 ArrayRef<uint64_t> Args) { 1496 // Evaluate each function and store the result in each target's RetVal 1497 // field. 1498 for (VirtualCallTarget &Target : TargetsForSlot) { 1499 if (Target.Fn->arg_size() != Args.size() + 1) 1500 return false; 1501 1502 Evaluator Eval(M.getDataLayout(), nullptr); 1503 SmallVector<Constant *, 2> EvalArgs; 1504 EvalArgs.push_back( 1505 Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0))); 1506 for (unsigned I = 0; I != Args.size(); ++I) { 1507 auto *ArgTy = dyn_cast<IntegerType>( 1508 Target.Fn->getFunctionType()->getParamType(I + 1)); 1509 if (!ArgTy) 1510 return false; 1511 EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I])); 1512 } 1513 1514 Constant *RetVal; 1515 if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) || 1516 !isa<ConstantInt>(RetVal)) 1517 return false; 1518 Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue(); 1519 } 1520 return true; 1521 } 1522 1523 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1524 uint64_t TheRetVal) { 1525 for (auto Call : CSInfo.CallSites) { 1526 if (!OptimizedCalls.insert(&Call.CB).second) 1527 continue; 1528 NumUniformRetVal++; 1529 Call.replaceAndErase( 1530 "uniform-ret-val", FnName, RemarksEnabled, OREGetter, 1531 ConstantInt::get(cast<IntegerType>(Call.CB.getType()), TheRetVal)); 1532 } 1533 CSInfo.markDevirt(); 1534 } 1535 1536 bool DevirtModule::tryUniformRetValOpt( 1537 MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo, 1538 WholeProgramDevirtResolution::ByArg *Res) { 1539 // Uniform return value optimization. If all functions return the same 1540 // constant, replace all calls with that constant. 1541 uint64_t TheRetVal = TargetsForSlot[0].RetVal; 1542 for (const VirtualCallTarget &Target : TargetsForSlot) 1543 if (Target.RetVal != TheRetVal) 1544 return false; 1545 1546 if (CSInfo.isExported()) { 1547 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniformRetVal; 1548 Res->Info = TheRetVal; 1549 } 1550 1551 applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal); 1552 if (RemarksEnabled || AreStatisticsEnabled()) 1553 for (auto &&Target : TargetsForSlot) 1554 Target.WasDevirt = true; 1555 return true; 1556 } 1557 1558 std::string DevirtModule::getGlobalName(VTableSlot Slot, 1559 ArrayRef<uint64_t> Args, 1560 StringRef Name) { 1561 std::string FullName = "__typeid_"; 1562 raw_string_ostream OS(FullName); 1563 OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset; 1564 for (uint64_t Arg : Args) 1565 OS << '_' << Arg; 1566 OS << '_' << Name; 1567 return OS.str(); 1568 } 1569 1570 bool DevirtModule::shouldExportConstantsAsAbsoluteSymbols() { 1571 Triple T(M.getTargetTriple()); 1572 return T.isX86() && T.getObjectFormat() == Triple::ELF; 1573 } 1574 1575 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1576 StringRef Name, Constant *C) { 1577 GlobalAlias *GA = GlobalAlias::create(Int8Ty, 0, GlobalValue::ExternalLinkage, 1578 getGlobalName(Slot, Args, Name), C, &M); 1579 GA->setVisibility(GlobalValue::HiddenVisibility); 1580 } 1581 1582 void DevirtModule::exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1583 StringRef Name, uint32_t Const, 1584 uint32_t &Storage) { 1585 if (shouldExportConstantsAsAbsoluteSymbols()) { 1586 exportGlobal( 1587 Slot, Args, Name, 1588 ConstantExpr::getIntToPtr(ConstantInt::get(Int32Ty, Const), Int8PtrTy)); 1589 return; 1590 } 1591 1592 Storage = Const; 1593 } 1594 1595 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1596 StringRef Name) { 1597 Constant *C = 1598 M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Arr0Ty); 1599 auto *GV = dyn_cast<GlobalVariable>(C); 1600 if (GV) 1601 GV->setVisibility(GlobalValue::HiddenVisibility); 1602 return C; 1603 } 1604 1605 Constant *DevirtModule::importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1606 StringRef Name, IntegerType *IntTy, 1607 uint32_t Storage) { 1608 if (!shouldExportConstantsAsAbsoluteSymbols()) 1609 return ConstantInt::get(IntTy, Storage); 1610 1611 Constant *C = importGlobal(Slot, Args, Name); 1612 auto *GV = cast<GlobalVariable>(C->stripPointerCasts()); 1613 C = ConstantExpr::getPtrToInt(C, IntTy); 1614 1615 // We only need to set metadata if the global is newly created, in which 1616 // case it would not have hidden visibility. 1617 if (GV->hasMetadata(LLVMContext::MD_absolute_symbol)) 1618 return C; 1619 1620 auto SetAbsRange = [&](uint64_t Min, uint64_t Max) { 1621 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min)); 1622 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max)); 1623 GV->setMetadata(LLVMContext::MD_absolute_symbol, 1624 MDNode::get(M.getContext(), {MinC, MaxC})); 1625 }; 1626 unsigned AbsWidth = IntTy->getBitWidth(); 1627 if (AbsWidth == IntPtrTy->getBitWidth()) 1628 SetAbsRange(~0ull, ~0ull); // Full set. 1629 else 1630 SetAbsRange(0, 1ull << AbsWidth); 1631 return C; 1632 } 1633 1634 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1635 bool IsOne, 1636 Constant *UniqueMemberAddr) { 1637 for (auto &&Call : CSInfo.CallSites) { 1638 if (!OptimizedCalls.insert(&Call.CB).second) 1639 continue; 1640 IRBuilder<> B(&Call.CB); 1641 Value *Cmp = 1642 B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE, Call.VTable, 1643 B.CreateBitCast(UniqueMemberAddr, Call.VTable->getType())); 1644 Cmp = B.CreateZExt(Cmp, Call.CB.getType()); 1645 NumUniqueRetVal++; 1646 Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter, 1647 Cmp); 1648 } 1649 CSInfo.markDevirt(); 1650 } 1651 1652 Constant *DevirtModule::getMemberAddr(const TypeMemberInfo *M) { 1653 Constant *C = ConstantExpr::getBitCast(M->Bits->GV, Int8PtrTy); 1654 return ConstantExpr::getGetElementPtr(Int8Ty, C, 1655 ConstantInt::get(Int64Ty, M->Offset)); 1656 } 1657 1658 bool DevirtModule::tryUniqueRetValOpt( 1659 unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1660 CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res, 1661 VTableSlot Slot, ArrayRef<uint64_t> Args) { 1662 // IsOne controls whether we look for a 0 or a 1. 1663 auto tryUniqueRetValOptFor = [&](bool IsOne) { 1664 const TypeMemberInfo *UniqueMember = nullptr; 1665 for (const VirtualCallTarget &Target : TargetsForSlot) { 1666 if (Target.RetVal == (IsOne ? 1 : 0)) { 1667 if (UniqueMember) 1668 return false; 1669 UniqueMember = Target.TM; 1670 } 1671 } 1672 1673 // We should have found a unique member or bailed out by now. We already 1674 // checked for a uniform return value in tryUniformRetValOpt. 1675 assert(UniqueMember); 1676 1677 Constant *UniqueMemberAddr = getMemberAddr(UniqueMember); 1678 if (CSInfo.isExported()) { 1679 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniqueRetVal; 1680 Res->Info = IsOne; 1681 1682 exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr); 1683 } 1684 1685 // Replace each call with the comparison. 1686 applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne, 1687 UniqueMemberAddr); 1688 1689 // Update devirtualization statistics for targets. 1690 if (RemarksEnabled || AreStatisticsEnabled()) 1691 for (auto &&Target : TargetsForSlot) 1692 Target.WasDevirt = true; 1693 1694 return true; 1695 }; 1696 1697 if (BitWidth == 1) { 1698 if (tryUniqueRetValOptFor(true)) 1699 return true; 1700 if (tryUniqueRetValOptFor(false)) 1701 return true; 1702 } 1703 return false; 1704 } 1705 1706 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 1707 Constant *Byte, Constant *Bit) { 1708 for (auto Call : CSInfo.CallSites) { 1709 if (!OptimizedCalls.insert(&Call.CB).second) 1710 continue; 1711 auto *RetType = cast<IntegerType>(Call.CB.getType()); 1712 IRBuilder<> B(&Call.CB); 1713 Value *Addr = 1714 B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte); 1715 if (RetType->getBitWidth() == 1) { 1716 Value *Bits = B.CreateLoad(Int8Ty, Addr); 1717 Value *BitsAndBit = B.CreateAnd(Bits, Bit); 1718 auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0)); 1719 NumVirtConstProp1Bit++; 1720 Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled, 1721 OREGetter, IsBitSet); 1722 } else { 1723 Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo()); 1724 Value *Val = B.CreateLoad(RetType, ValAddr); 1725 NumVirtConstProp++; 1726 Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled, 1727 OREGetter, Val); 1728 } 1729 } 1730 CSInfo.markDevirt(); 1731 } 1732 1733 bool DevirtModule::tryVirtualConstProp( 1734 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1735 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1736 // This only works if the function returns an integer. 1737 auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType()); 1738 if (!RetType) 1739 return false; 1740 unsigned BitWidth = RetType->getBitWidth(); 1741 if (BitWidth > 64) 1742 return false; 1743 1744 // Make sure that each function is defined, does not access memory, takes at 1745 // least one argument, does not use its first argument (which we assume is 1746 // 'this'), and has the same return type. 1747 // 1748 // Note that we test whether this copy of the function is readnone, rather 1749 // than testing function attributes, which must hold for any copy of the 1750 // function, even a less optimized version substituted at link time. This is 1751 // sound because the virtual constant propagation optimizations effectively 1752 // inline all implementations of the virtual function into each call site, 1753 // rather than using function attributes to perform local optimization. 1754 for (VirtualCallTarget &Target : TargetsForSlot) { 1755 if (Target.Fn->isDeclaration() || 1756 computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) != 1757 FMRB_DoesNotAccessMemory || 1758 Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() || 1759 Target.Fn->getReturnType() != RetType) 1760 return false; 1761 } 1762 1763 for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) { 1764 if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first)) 1765 continue; 1766 1767 WholeProgramDevirtResolution::ByArg *ResByArg = nullptr; 1768 if (Res) 1769 ResByArg = &Res->ResByArg[CSByConstantArg.first]; 1770 1771 if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg)) 1772 continue; 1773 1774 if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second, 1775 ResByArg, Slot, CSByConstantArg.first)) 1776 continue; 1777 1778 // Find an allocation offset in bits in all vtables associated with the 1779 // type. 1780 uint64_t AllocBefore = 1781 findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth); 1782 uint64_t AllocAfter = 1783 findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth); 1784 1785 // Calculate the total amount of padding needed to store a value at both 1786 // ends of the object. 1787 uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0; 1788 for (auto &&Target : TargetsForSlot) { 1789 TotalPaddingBefore += std::max<int64_t>( 1790 (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0); 1791 TotalPaddingAfter += std::max<int64_t>( 1792 (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0); 1793 } 1794 1795 // If the amount of padding is too large, give up. 1796 // FIXME: do something smarter here. 1797 if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128) 1798 continue; 1799 1800 // Calculate the offset to the value as a (possibly negative) byte offset 1801 // and (if applicable) a bit offset, and store the values in the targets. 1802 int64_t OffsetByte; 1803 uint64_t OffsetBit; 1804 if (TotalPaddingBefore <= TotalPaddingAfter) 1805 setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte, 1806 OffsetBit); 1807 else 1808 setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte, 1809 OffsetBit); 1810 1811 if (RemarksEnabled || AreStatisticsEnabled()) 1812 for (auto &&Target : TargetsForSlot) 1813 Target.WasDevirt = true; 1814 1815 1816 if (CSByConstantArg.second.isExported()) { 1817 ResByArg->TheKind = WholeProgramDevirtResolution::ByArg::VirtualConstProp; 1818 exportConstant(Slot, CSByConstantArg.first, "byte", OffsetByte, 1819 ResByArg->Byte); 1820 exportConstant(Slot, CSByConstantArg.first, "bit", 1ULL << OffsetBit, 1821 ResByArg->Bit); 1822 } 1823 1824 // Rewrite each call to a load from OffsetByte/OffsetBit. 1825 Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte); 1826 Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit); 1827 applyVirtualConstProp(CSByConstantArg.second, 1828 TargetsForSlot[0].Fn->getName(), ByteConst, BitConst); 1829 } 1830 return true; 1831 } 1832 1833 void DevirtModule::rebuildGlobal(VTableBits &B) { 1834 if (B.Before.Bytes.empty() && B.After.Bytes.empty()) 1835 return; 1836 1837 // Align the before byte array to the global's minimum alignment so that we 1838 // don't break any alignment requirements on the global. 1839 Align Alignment = M.getDataLayout().getValueOrABITypeAlignment( 1840 B.GV->getAlign(), B.GV->getValueType()); 1841 B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), Alignment)); 1842 1843 // Before was stored in reverse order; flip it now. 1844 for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I) 1845 std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]); 1846 1847 // Build an anonymous global containing the before bytes, followed by the 1848 // original initializer, followed by the after bytes. 1849 auto NewInit = ConstantStruct::getAnon( 1850 {ConstantDataArray::get(M.getContext(), B.Before.Bytes), 1851 B.GV->getInitializer(), 1852 ConstantDataArray::get(M.getContext(), B.After.Bytes)}); 1853 auto NewGV = 1854 new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(), 1855 GlobalVariable::PrivateLinkage, NewInit, "", B.GV); 1856 NewGV->setSection(B.GV->getSection()); 1857 NewGV->setComdat(B.GV->getComdat()); 1858 NewGV->setAlignment(B.GV->getAlign()); 1859 1860 // Copy the original vtable's metadata to the anonymous global, adjusting 1861 // offsets as required. 1862 NewGV->copyMetadata(B.GV, B.Before.Bytes.size()); 1863 1864 // Build an alias named after the original global, pointing at the second 1865 // element (the original initializer). 1866 auto Alias = GlobalAlias::create( 1867 B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "", 1868 ConstantExpr::getGetElementPtr( 1869 NewInit->getType(), NewGV, 1870 ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0), 1871 ConstantInt::get(Int32Ty, 1)}), 1872 &M); 1873 Alias->setVisibility(B.GV->getVisibility()); 1874 Alias->takeName(B.GV); 1875 1876 B.GV->replaceAllUsesWith(Alias); 1877 B.GV->eraseFromParent(); 1878 } 1879 1880 bool DevirtModule::areRemarksEnabled() { 1881 const auto &FL = M.getFunctionList(); 1882 for (const Function &Fn : FL) { 1883 const auto &BBL = Fn.getBasicBlockList(); 1884 if (BBL.empty()) 1885 continue; 1886 auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front()); 1887 return DI.isEnabled(); 1888 } 1889 return false; 1890 } 1891 1892 void DevirtModule::scanTypeTestUsers( 1893 Function *TypeTestFunc, 1894 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 1895 // Find all virtual calls via a virtual table pointer %p under an assumption 1896 // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p 1897 // points to a member of the type identifier %md. Group calls by (type ID, 1898 // offset) pair (effectively the identity of the virtual function) and store 1899 // to CallSlots. 1900 for (Use &U : llvm::make_early_inc_range(TypeTestFunc->uses())) { 1901 auto *CI = dyn_cast<CallInst>(U.getUser()); 1902 if (!CI) 1903 continue; 1904 1905 // Search for virtual calls based on %p and add them to DevirtCalls. 1906 SmallVector<DevirtCallSite, 1> DevirtCalls; 1907 SmallVector<CallInst *, 1> Assumes; 1908 auto &DT = LookupDomTree(*CI->getFunction()); 1909 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT); 1910 1911 Metadata *TypeId = 1912 cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata(); 1913 // If we found any, add them to CallSlots. 1914 if (!Assumes.empty()) { 1915 Value *Ptr = CI->getArgOperand(0)->stripPointerCasts(); 1916 for (DevirtCallSite Call : DevirtCalls) 1917 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, nullptr); 1918 } 1919 1920 auto RemoveTypeTestAssumes = [&]() { 1921 // We no longer need the assumes or the type test. 1922 for (auto Assume : Assumes) 1923 Assume->eraseFromParent(); 1924 // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we 1925 // may use the vtable argument later. 1926 if (CI->use_empty()) 1927 CI->eraseFromParent(); 1928 }; 1929 1930 // At this point we could remove all type test assume sequences, as they 1931 // were originally inserted for WPD. However, we can keep these in the 1932 // code stream for later analysis (e.g. to help drive more efficient ICP 1933 // sequences). They will eventually be removed by a second LowerTypeTests 1934 // invocation that cleans them up. In order to do this correctly, the first 1935 // LowerTypeTests invocation needs to know that they have "Unknown" type 1936 // test resolution, so that they aren't treated as Unsat and lowered to 1937 // False, which will break any uses on assumes. Below we remove any type 1938 // test assumes that will not be treated as Unknown by LTT. 1939 1940 // The type test assumes will be treated by LTT as Unsat if the type id is 1941 // not used on a global (in which case it has no entry in the TypeIdMap). 1942 if (!TypeIdMap.count(TypeId)) 1943 RemoveTypeTestAssumes(); 1944 1945 // For ThinLTO importing, we need to remove the type test assumes if this is 1946 // an MDString type id without a corresponding TypeIdSummary. Any 1947 // non-MDString type ids are ignored and treated as Unknown by LTT, so their 1948 // type test assumes can be kept. If the MDString type id is missing a 1949 // TypeIdSummary (e.g. because there was no use on a vcall, preventing the 1950 // exporting phase of WPD from analyzing it), then it would be treated as 1951 // Unsat by LTT and we need to remove its type test assumes here. If not 1952 // used on a vcall we don't need them for later optimization use in any 1953 // case. 1954 else if (ImportSummary && isa<MDString>(TypeId)) { 1955 const TypeIdSummary *TidSummary = 1956 ImportSummary->getTypeIdSummary(cast<MDString>(TypeId)->getString()); 1957 if (!TidSummary) 1958 RemoveTypeTestAssumes(); 1959 else 1960 // If one was created it should not be Unsat, because if we reached here 1961 // the type id was used on a global. 1962 assert(TidSummary->TTRes.TheKind != TypeTestResolution::Unsat); 1963 } 1964 } 1965 } 1966 1967 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) { 1968 Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test); 1969 1970 for (Use &U : llvm::make_early_inc_range(TypeCheckedLoadFunc->uses())) { 1971 auto *CI = dyn_cast<CallInst>(U.getUser()); 1972 if (!CI) 1973 continue; 1974 1975 Value *Ptr = CI->getArgOperand(0); 1976 Value *Offset = CI->getArgOperand(1); 1977 Value *TypeIdValue = CI->getArgOperand(2); 1978 Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata(); 1979 1980 SmallVector<DevirtCallSite, 1> DevirtCalls; 1981 SmallVector<Instruction *, 1> LoadedPtrs; 1982 SmallVector<Instruction *, 1> Preds; 1983 bool HasNonCallUses = false; 1984 auto &DT = LookupDomTree(*CI->getFunction()); 1985 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds, 1986 HasNonCallUses, CI, DT); 1987 1988 // Start by generating "pessimistic" code that explicitly loads the function 1989 // pointer from the vtable and performs the type check. If possible, we will 1990 // eliminate the load and the type check later. 1991 1992 // If possible, only generate the load at the point where it is used. 1993 // This helps avoid unnecessary spills. 1994 IRBuilder<> LoadB( 1995 (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI); 1996 Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset); 1997 Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy)); 1998 Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr); 1999 2000 for (Instruction *LoadedPtr : LoadedPtrs) { 2001 LoadedPtr->replaceAllUsesWith(LoadedValue); 2002 LoadedPtr->eraseFromParent(); 2003 } 2004 2005 // Likewise for the type test. 2006 IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI); 2007 CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue}); 2008 2009 for (Instruction *Pred : Preds) { 2010 Pred->replaceAllUsesWith(TypeTestCall); 2011 Pred->eraseFromParent(); 2012 } 2013 2014 // We have already erased any extractvalue instructions that refer to the 2015 // intrinsic call, but the intrinsic may have other non-extractvalue uses 2016 // (although this is unlikely). In that case, explicitly build a pair and 2017 // RAUW it. 2018 if (!CI->use_empty()) { 2019 Value *Pair = UndefValue::get(CI->getType()); 2020 IRBuilder<> B(CI); 2021 Pair = B.CreateInsertValue(Pair, LoadedValue, {0}); 2022 Pair = B.CreateInsertValue(Pair, TypeTestCall, {1}); 2023 CI->replaceAllUsesWith(Pair); 2024 } 2025 2026 // The number of unsafe uses is initially the number of uses. 2027 auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall]; 2028 NumUnsafeUses = DevirtCalls.size(); 2029 2030 // If the function pointer has a non-call user, we cannot eliminate the type 2031 // check, as one of those users may eventually call the pointer. Increment 2032 // the unsafe use count to make sure it cannot reach zero. 2033 if (HasNonCallUses) 2034 ++NumUnsafeUses; 2035 for (DevirtCallSite Call : DevirtCalls) { 2036 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, 2037 &NumUnsafeUses); 2038 } 2039 2040 CI->eraseFromParent(); 2041 } 2042 } 2043 2044 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) { 2045 auto *TypeId = dyn_cast<MDString>(Slot.TypeID); 2046 if (!TypeId) 2047 return; 2048 const TypeIdSummary *TidSummary = 2049 ImportSummary->getTypeIdSummary(TypeId->getString()); 2050 if (!TidSummary) 2051 return; 2052 auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset); 2053 if (ResI == TidSummary->WPDRes.end()) 2054 return; 2055 const WholeProgramDevirtResolution &Res = ResI->second; 2056 2057 if (Res.TheKind == WholeProgramDevirtResolution::SingleImpl) { 2058 assert(!Res.SingleImplName.empty()); 2059 // The type of the function in the declaration is irrelevant because every 2060 // call site will cast it to the correct type. 2061 Constant *SingleImpl = 2062 cast<Constant>(M.getOrInsertFunction(Res.SingleImplName, 2063 Type::getVoidTy(M.getContext())) 2064 .getCallee()); 2065 2066 // This is the import phase so we should not be exporting anything. 2067 bool IsExported = false; 2068 applySingleImplDevirt(SlotInfo, SingleImpl, IsExported); 2069 assert(!IsExported); 2070 } 2071 2072 for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) { 2073 auto I = Res.ResByArg.find(CSByConstantArg.first); 2074 if (I == Res.ResByArg.end()) 2075 continue; 2076 auto &ResByArg = I->second; 2077 // FIXME: We should figure out what to do about the "function name" argument 2078 // to the apply* functions, as the function names are unavailable during the 2079 // importing phase. For now we just pass the empty string. This does not 2080 // impact correctness because the function names are just used for remarks. 2081 switch (ResByArg.TheKind) { 2082 case WholeProgramDevirtResolution::ByArg::UniformRetVal: 2083 applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info); 2084 break; 2085 case WholeProgramDevirtResolution::ByArg::UniqueRetVal: { 2086 Constant *UniqueMemberAddr = 2087 importGlobal(Slot, CSByConstantArg.first, "unique_member"); 2088 applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info, 2089 UniqueMemberAddr); 2090 break; 2091 } 2092 case WholeProgramDevirtResolution::ByArg::VirtualConstProp: { 2093 Constant *Byte = importConstant(Slot, CSByConstantArg.first, "byte", 2094 Int32Ty, ResByArg.Byte); 2095 Constant *Bit = importConstant(Slot, CSByConstantArg.first, "bit", Int8Ty, 2096 ResByArg.Bit); 2097 applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit); 2098 break; 2099 } 2100 default: 2101 break; 2102 } 2103 } 2104 2105 if (Res.TheKind == WholeProgramDevirtResolution::BranchFunnel) { 2106 // The type of the function is irrelevant, because it's bitcast at calls 2107 // anyhow. 2108 Constant *JT = cast<Constant>( 2109 M.getOrInsertFunction(getGlobalName(Slot, {}, "branch_funnel"), 2110 Type::getVoidTy(M.getContext())) 2111 .getCallee()); 2112 bool IsExported = false; 2113 applyICallBranchFunnel(SlotInfo, JT, IsExported); 2114 assert(!IsExported); 2115 } 2116 } 2117 2118 void DevirtModule::removeRedundantTypeTests() { 2119 auto True = ConstantInt::getTrue(M.getContext()); 2120 for (auto &&U : NumUnsafeUsesForTypeTest) { 2121 if (U.second == 0) { 2122 U.first->replaceAllUsesWith(True); 2123 U.first->eraseFromParent(); 2124 } 2125 } 2126 } 2127 2128 ValueInfo 2129 DevirtModule::lookUpFunctionValueInfo(Function *TheFn, 2130 ModuleSummaryIndex *ExportSummary) { 2131 assert((ExportSummary != nullptr) && 2132 "Caller guarantees ExportSummary is not nullptr"); 2133 2134 const auto TheFnGUID = TheFn->getGUID(); 2135 const auto TheFnGUIDWithExportedName = GlobalValue::getGUID(TheFn->getName()); 2136 // Look up ValueInfo with the GUID in the current linkage. 2137 ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFnGUID); 2138 // If no entry is found and GUID is different from GUID computed using 2139 // exported name, look up ValueInfo with the exported name unconditionally. 2140 // This is a fallback. 2141 // 2142 // The reason to have a fallback: 2143 // 1. LTO could enable global value internalization via 2144 // `enable-lto-internalization`. 2145 // 2. The GUID in ExportedSummary is computed using exported name. 2146 if ((!TheFnVI) && (TheFnGUID != TheFnGUIDWithExportedName)) { 2147 TheFnVI = ExportSummary->getValueInfo(TheFnGUIDWithExportedName); 2148 } 2149 return TheFnVI; 2150 } 2151 2152 bool DevirtModule::mustBeUnreachableFunction( 2153 Function *const F, ModuleSummaryIndex *ExportSummary) { 2154 // First, learn unreachability by analyzing function IR. 2155 if (!F->isDeclaration()) { 2156 // A function must be unreachable if its entry block ends with an 2157 // 'unreachable'. 2158 return isa<UnreachableInst>(F->getEntryBlock().getTerminator()); 2159 } 2160 // Learn unreachability from ExportSummary if ExportSummary is present. 2161 return ExportSummary && 2162 ::mustBeUnreachableFunction( 2163 DevirtModule::lookUpFunctionValueInfo(F, ExportSummary)); 2164 } 2165 2166 bool DevirtModule::run() { 2167 // If only some of the modules were split, we cannot correctly perform 2168 // this transformation. We already checked for the presense of type tests 2169 // with partially split modules during the thin link, and would have emitted 2170 // an error if any were found, so here we can simply return. 2171 if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) || 2172 (ImportSummary && ImportSummary->partiallySplitLTOUnits())) 2173 return false; 2174 2175 Function *TypeTestFunc = 2176 M.getFunction(Intrinsic::getName(Intrinsic::type_test)); 2177 Function *TypeCheckedLoadFunc = 2178 M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load)); 2179 Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume)); 2180 2181 // Normally if there are no users of the devirtualization intrinsics in the 2182 // module, this pass has nothing to do. But if we are exporting, we also need 2183 // to handle any users that appear only in the function summaries. 2184 if (!ExportSummary && 2185 (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc || 2186 AssumeFunc->use_empty()) && 2187 (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty())) 2188 return false; 2189 2190 // Rebuild type metadata into a map for easy lookup. 2191 std::vector<VTableBits> Bits; 2192 DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap; 2193 buildTypeIdentifierMap(Bits, TypeIdMap); 2194 2195 if (TypeTestFunc && AssumeFunc) 2196 scanTypeTestUsers(TypeTestFunc, TypeIdMap); 2197 2198 if (TypeCheckedLoadFunc) 2199 scanTypeCheckedLoadUsers(TypeCheckedLoadFunc); 2200 2201 if (ImportSummary) { 2202 for (auto &S : CallSlots) 2203 importResolution(S.first, S.second); 2204 2205 removeRedundantTypeTests(); 2206 2207 // We have lowered or deleted the type intrinsics, so we will no longer have 2208 // enough information to reason about the liveness of virtual function 2209 // pointers in GlobalDCE. 2210 for (GlobalVariable &GV : M.globals()) 2211 GV.eraseMetadata(LLVMContext::MD_vcall_visibility); 2212 2213 // The rest of the code is only necessary when exporting or during regular 2214 // LTO, so we are done. 2215 return true; 2216 } 2217 2218 if (TypeIdMap.empty()) 2219 return true; 2220 2221 // Collect information from summary about which calls to try to devirtualize. 2222 if (ExportSummary) { 2223 DenseMap<GlobalValue::GUID, TinyPtrVector<Metadata *>> MetadataByGUID; 2224 for (auto &P : TypeIdMap) { 2225 if (auto *TypeId = dyn_cast<MDString>(P.first)) 2226 MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back( 2227 TypeId); 2228 } 2229 2230 for (auto &P : *ExportSummary) { 2231 for (auto &S : P.second.SummaryList) { 2232 auto *FS = dyn_cast<FunctionSummary>(S.get()); 2233 if (!FS) 2234 continue; 2235 // FIXME: Only add live functions. 2236 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 2237 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 2238 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 2239 } 2240 } 2241 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 2242 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 2243 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 2244 } 2245 } 2246 for (const FunctionSummary::ConstVCall &VC : 2247 FS->type_test_assume_const_vcalls()) { 2248 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 2249 CallSlots[{MD, VC.VFunc.Offset}] 2250 .ConstCSInfo[VC.Args] 2251 .addSummaryTypeTestAssumeUser(FS); 2252 } 2253 } 2254 for (const FunctionSummary::ConstVCall &VC : 2255 FS->type_checked_load_const_vcalls()) { 2256 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 2257 CallSlots[{MD, VC.VFunc.Offset}] 2258 .ConstCSInfo[VC.Args] 2259 .addSummaryTypeCheckedLoadUser(FS); 2260 } 2261 } 2262 } 2263 } 2264 } 2265 2266 // For each (type, offset) pair: 2267 bool DidVirtualConstProp = false; 2268 std::map<std::string, Function*> DevirtTargets; 2269 for (auto &S : CallSlots) { 2270 // Search each of the members of the type identifier for the virtual 2271 // function implementation at offset S.first.ByteOffset, and add to 2272 // TargetsForSlot. 2273 std::vector<VirtualCallTarget> TargetsForSlot; 2274 WholeProgramDevirtResolution *Res = nullptr; 2275 const std::set<TypeMemberInfo> &TypeMemberInfos = TypeIdMap[S.first.TypeID]; 2276 if (ExportSummary && isa<MDString>(S.first.TypeID) && 2277 TypeMemberInfos.size()) 2278 // For any type id used on a global's type metadata, create the type id 2279 // summary resolution regardless of whether we can devirtualize, so that 2280 // lower type tests knows the type id is not Unsat. If it was not used on 2281 // a global's type metadata, the TypeIdMap entry set will be empty, and 2282 // we don't want to create an entry (with the default Unknown type 2283 // resolution), which can prevent detection of the Unsat. 2284 Res = &ExportSummary 2285 ->getOrInsertTypeIdSummary( 2286 cast<MDString>(S.first.TypeID)->getString()) 2287 .WPDRes[S.first.ByteOffset]; 2288 if (tryFindVirtualCallTargets(TargetsForSlot, TypeMemberInfos, 2289 S.first.ByteOffset, ExportSummary)) { 2290 2291 if (!trySingleImplDevirt(ExportSummary, TargetsForSlot, S.second, Res)) { 2292 DidVirtualConstProp |= 2293 tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first); 2294 2295 tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first); 2296 } 2297 2298 // Collect functions devirtualized at least for one call site for stats. 2299 if (RemarksEnabled || AreStatisticsEnabled()) 2300 for (const auto &T : TargetsForSlot) 2301 if (T.WasDevirt) 2302 DevirtTargets[std::string(T.Fn->getName())] = T.Fn; 2303 } 2304 2305 // CFI-specific: if we are exporting and any llvm.type.checked.load 2306 // intrinsics were *not* devirtualized, we need to add the resulting 2307 // llvm.type.test intrinsics to the function summaries so that the 2308 // LowerTypeTests pass will export them. 2309 if (ExportSummary && isa<MDString>(S.first.TypeID)) { 2310 auto GUID = 2311 GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString()); 2312 for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers) 2313 FS->addTypeTest(GUID); 2314 for (auto &CCS : S.second.ConstCSInfo) 2315 for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers) 2316 FS->addTypeTest(GUID); 2317 } 2318 } 2319 2320 if (RemarksEnabled) { 2321 // Generate remarks for each devirtualized function. 2322 for (const auto &DT : DevirtTargets) { 2323 Function *F = DT.second; 2324 2325 using namespace ore; 2326 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F) 2327 << "devirtualized " 2328 << NV("FunctionName", DT.first)); 2329 } 2330 } 2331 2332 NumDevirtTargets += DevirtTargets.size(); 2333 2334 removeRedundantTypeTests(); 2335 2336 // Rebuild each global we touched as part of virtual constant propagation to 2337 // include the before and after bytes. 2338 if (DidVirtualConstProp) 2339 for (VTableBits &B : Bits) 2340 rebuildGlobal(B); 2341 2342 // We have lowered or deleted the type intrinsics, so we will no longer have 2343 // enough information to reason about the liveness of virtual function 2344 // pointers in GlobalDCE. 2345 for (GlobalVariable &GV : M.globals()) 2346 GV.eraseMetadata(LLVMContext::MD_vcall_visibility); 2347 2348 return true; 2349 } 2350 2351 void DevirtIndex::run() { 2352 if (ExportSummary.typeIdCompatibleVtableMap().empty()) 2353 return; 2354 2355 DenseMap<GlobalValue::GUID, std::vector<StringRef>> NameByGUID; 2356 for (auto &P : ExportSummary.typeIdCompatibleVtableMap()) { 2357 NameByGUID[GlobalValue::getGUID(P.first)].push_back(P.first); 2358 } 2359 2360 // Collect information from summary about which calls to try to devirtualize. 2361 for (auto &P : ExportSummary) { 2362 for (auto &S : P.second.SummaryList) { 2363 auto *FS = dyn_cast<FunctionSummary>(S.get()); 2364 if (!FS) 2365 continue; 2366 // FIXME: Only add live functions. 2367 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 2368 for (StringRef Name : NameByGUID[VF.GUID]) { 2369 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 2370 } 2371 } 2372 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 2373 for (StringRef Name : NameByGUID[VF.GUID]) { 2374 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 2375 } 2376 } 2377 for (const FunctionSummary::ConstVCall &VC : 2378 FS->type_test_assume_const_vcalls()) { 2379 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 2380 CallSlots[{Name, VC.VFunc.Offset}] 2381 .ConstCSInfo[VC.Args] 2382 .addSummaryTypeTestAssumeUser(FS); 2383 } 2384 } 2385 for (const FunctionSummary::ConstVCall &VC : 2386 FS->type_checked_load_const_vcalls()) { 2387 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 2388 CallSlots[{Name, VC.VFunc.Offset}] 2389 .ConstCSInfo[VC.Args] 2390 .addSummaryTypeCheckedLoadUser(FS); 2391 } 2392 } 2393 } 2394 } 2395 2396 std::set<ValueInfo> DevirtTargets; 2397 // For each (type, offset) pair: 2398 for (auto &S : CallSlots) { 2399 // Search each of the members of the type identifier for the virtual 2400 // function implementation at offset S.first.ByteOffset, and add to 2401 // TargetsForSlot. 2402 std::vector<ValueInfo> TargetsForSlot; 2403 auto TidSummary = ExportSummary.getTypeIdCompatibleVtableSummary(S.first.TypeID); 2404 assert(TidSummary); 2405 // Create the type id summary resolution regardlness of whether we can 2406 // devirtualize, so that lower type tests knows the type id is used on 2407 // a global and not Unsat. 2408 WholeProgramDevirtResolution *Res = 2409 &ExportSummary.getOrInsertTypeIdSummary(S.first.TypeID) 2410 .WPDRes[S.first.ByteOffset]; 2411 if (tryFindVirtualCallTargets(TargetsForSlot, *TidSummary, 2412 S.first.ByteOffset)) { 2413 2414 if (!trySingleImplDevirt(TargetsForSlot, S.first, S.second, Res, 2415 DevirtTargets)) 2416 continue; 2417 } 2418 } 2419 2420 // Optionally have the thin link print message for each devirtualized 2421 // function. 2422 if (PrintSummaryDevirt) 2423 for (const auto &DT : DevirtTargets) 2424 errs() << "Devirtualized call to " << DT << "\n"; 2425 2426 NumDevirtTargets += DevirtTargets.size(); 2427 } 2428