1 //===- WholeProgramDevirt.cpp - Whole program virtual call optimization ---===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This pass implements whole program optimization of virtual calls in cases 10 // where we know (via !type metadata) that the list of callees is fixed. This 11 // includes the following: 12 // - Single implementation devirtualization: if a virtual call has a single 13 // possible callee, replace all calls with a direct call to that callee. 14 // - Virtual constant propagation: if the virtual function's return type is an 15 // integer <=64 bits and all possible callees are readnone, for each class and 16 // each list of constant arguments: evaluate the function, store the return 17 // value alongside the virtual table, and rewrite each virtual call as a load 18 // from the virtual table. 19 // - Uniform return value optimization: if the conditions for virtual constant 20 // propagation hold and each function returns the same constant value, replace 21 // each virtual call with that constant. 22 // - Unique return value optimization for i1 return values: if the conditions 23 // for virtual constant propagation hold and a single vtable's function 24 // returns 0, or a single vtable's function returns 1, replace each virtual 25 // call with a comparison of the vptr against that vtable's address. 26 // 27 // This pass is intended to be used during the regular and thin LTO pipelines: 28 // 29 // During regular LTO, the pass determines the best optimization for each 30 // virtual call and applies the resolutions directly to virtual calls that are 31 // eligible for virtual call optimization (i.e. calls that use either of the 32 // llvm.assume(llvm.type.test) or llvm.type.checked.load intrinsics). 33 // 34 // During hybrid Regular/ThinLTO, the pass operates in two phases: 35 // - Export phase: this is run during the thin link over a single merged module 36 // that contains all vtables with !type metadata that participate in the link. 37 // The pass computes a resolution for each virtual call and stores it in the 38 // type identifier summary. 39 // - Import phase: this is run during the thin backends over the individual 40 // modules. The pass applies the resolutions previously computed during the 41 // import phase to each eligible virtual call. 42 // 43 // During ThinLTO, the pass operates in two phases: 44 // - Export phase: this is run during the thin link over the index which 45 // contains a summary of all vtables with !type metadata that participate in 46 // the link. It computes a resolution for each virtual call and stores it in 47 // the type identifier summary. Only single implementation devirtualization 48 // is supported. 49 // - Import phase: (same as with hybrid case above). 50 // 51 //===----------------------------------------------------------------------===// 52 53 #include "llvm/Transforms/IPO/WholeProgramDevirt.h" 54 #include "llvm/ADT/ArrayRef.h" 55 #include "llvm/ADT/DenseMap.h" 56 #include "llvm/ADT/DenseMapInfo.h" 57 #include "llvm/ADT/DenseSet.h" 58 #include "llvm/ADT/MapVector.h" 59 #include "llvm/ADT/SmallVector.h" 60 #include "llvm/ADT/Statistic.h" 61 #include "llvm/ADT/Triple.h" 62 #include "llvm/ADT/iterator_range.h" 63 #include "llvm/Analysis/AssumptionCache.h" 64 #include "llvm/Analysis/BasicAliasAnalysis.h" 65 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 66 #include "llvm/Analysis/TypeMetadataUtils.h" 67 #include "llvm/Bitcode/BitcodeReader.h" 68 #include "llvm/Bitcode/BitcodeWriter.h" 69 #include "llvm/IR/Constants.h" 70 #include "llvm/IR/DataLayout.h" 71 #include "llvm/IR/DebugLoc.h" 72 #include "llvm/IR/DerivedTypes.h" 73 #include "llvm/IR/Dominators.h" 74 #include "llvm/IR/Function.h" 75 #include "llvm/IR/GlobalAlias.h" 76 #include "llvm/IR/GlobalVariable.h" 77 #include "llvm/IR/IRBuilder.h" 78 #include "llvm/IR/InstrTypes.h" 79 #include "llvm/IR/Instruction.h" 80 #include "llvm/IR/Instructions.h" 81 #include "llvm/IR/Intrinsics.h" 82 #include "llvm/IR/LLVMContext.h" 83 #include "llvm/IR/MDBuilder.h" 84 #include "llvm/IR/Metadata.h" 85 #include "llvm/IR/Module.h" 86 #include "llvm/IR/ModuleSummaryIndexYAML.h" 87 #include "llvm/InitializePasses.h" 88 #include "llvm/Pass.h" 89 #include "llvm/PassRegistry.h" 90 #include "llvm/Support/Casting.h" 91 #include "llvm/Support/CommandLine.h" 92 #include "llvm/Support/Errc.h" 93 #include "llvm/Support/Error.h" 94 #include "llvm/Support/FileSystem.h" 95 #include "llvm/Support/GlobPattern.h" 96 #include "llvm/Support/MathExtras.h" 97 #include "llvm/Transforms/IPO.h" 98 #include "llvm/Transforms/IPO/FunctionAttrs.h" 99 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 100 #include "llvm/Transforms/Utils/CallPromotionUtils.h" 101 #include "llvm/Transforms/Utils/Evaluator.h" 102 #include <algorithm> 103 #include <cstddef> 104 #include <map> 105 #include <set> 106 #include <string> 107 108 using namespace llvm; 109 using namespace wholeprogramdevirt; 110 111 #define DEBUG_TYPE "wholeprogramdevirt" 112 113 STATISTIC(NumDevirtTargets, "Number of whole program devirtualization targets"); 114 STATISTIC(NumSingleImpl, "Number of single implementation devirtualizations"); 115 STATISTIC(NumBranchFunnel, "Number of branch funnels"); 116 STATISTIC(NumUniformRetVal, "Number of uniform return value optimizations"); 117 STATISTIC(NumUniqueRetVal, "Number of unique return value optimizations"); 118 STATISTIC(NumVirtConstProp1Bit, 119 "Number of 1 bit virtual constant propagations"); 120 STATISTIC(NumVirtConstProp, "Number of virtual constant propagations"); 121 122 static cl::opt<PassSummaryAction> ClSummaryAction( 123 "wholeprogramdevirt-summary-action", 124 cl::desc("What to do with the summary when running this pass"), 125 cl::values(clEnumValN(PassSummaryAction::None, "none", "Do nothing"), 126 clEnumValN(PassSummaryAction::Import, "import", 127 "Import typeid resolutions from summary and globals"), 128 clEnumValN(PassSummaryAction::Export, "export", 129 "Export typeid resolutions to summary and globals")), 130 cl::Hidden); 131 132 static cl::opt<std::string> ClReadSummary( 133 "wholeprogramdevirt-read-summary", 134 cl::desc( 135 "Read summary from given bitcode or YAML file before running pass"), 136 cl::Hidden); 137 138 static cl::opt<std::string> ClWriteSummary( 139 "wholeprogramdevirt-write-summary", 140 cl::desc("Write summary to given bitcode or YAML file after running pass. " 141 "Output file format is deduced from extension: *.bc means writing " 142 "bitcode, otherwise YAML"), 143 cl::Hidden); 144 145 static cl::opt<unsigned> 146 ClThreshold("wholeprogramdevirt-branch-funnel-threshold", cl::Hidden, 147 cl::init(10), 148 cl::desc("Maximum number of call targets per " 149 "call site to enable branch funnels")); 150 151 static cl::opt<bool> 152 PrintSummaryDevirt("wholeprogramdevirt-print-index-based", cl::Hidden, 153 cl::desc("Print index-based devirtualization messages")); 154 155 /// Provide a way to force enable whole program visibility in tests. 156 /// This is needed to support legacy tests that don't contain 157 /// !vcall_visibility metadata (the mere presense of type tests 158 /// previously implied hidden visibility). 159 static cl::opt<bool> 160 WholeProgramVisibility("whole-program-visibility", cl::Hidden, 161 cl::desc("Enable whole program visibility")); 162 163 /// Provide a way to force disable whole program for debugging or workarounds, 164 /// when enabled via the linker. 165 static cl::opt<bool> DisableWholeProgramVisibility( 166 "disable-whole-program-visibility", cl::Hidden, 167 cl::desc("Disable whole program visibility (overrides enabling options)")); 168 169 /// Provide way to prevent certain function from being devirtualized 170 static cl::list<std::string> 171 SkipFunctionNames("wholeprogramdevirt-skip", 172 cl::desc("Prevent function(s) from being devirtualized"), 173 cl::Hidden, cl::CommaSeparated); 174 175 /// Mechanism to add runtime checking of devirtualization decisions, optionally 176 /// trapping or falling back to indirect call on any that are not correct. 177 /// Trapping mode is useful for debugging undefined behavior leading to failures 178 /// with WPD. Fallback mode is useful for ensuring safety when whole program 179 /// visibility may be compromised. 180 enum WPDCheckMode { None, Trap, Fallback }; 181 static cl::opt<WPDCheckMode> DevirtCheckMode( 182 "wholeprogramdevirt-check", cl::Hidden, 183 cl::desc("Type of checking for incorrect devirtualizations"), 184 cl::values(clEnumValN(WPDCheckMode::None, "none", "No checking"), 185 clEnumValN(WPDCheckMode::Trap, "trap", "Trap when incorrect"), 186 clEnumValN(WPDCheckMode::Fallback, "fallback", 187 "Fallback to indirect when incorrect"))); 188 189 namespace { 190 struct PatternList { 191 std::vector<GlobPattern> Patterns; 192 template <class T> void init(const T &StringList) { 193 for (const auto &S : StringList) 194 if (Expected<GlobPattern> Pat = GlobPattern::create(S)) 195 Patterns.push_back(std::move(*Pat)); 196 } 197 bool match(StringRef S) { 198 for (const GlobPattern &P : Patterns) 199 if (P.match(S)) 200 return true; 201 return false; 202 } 203 }; 204 } // namespace 205 206 // Find the minimum offset that we may store a value of size Size bits at. If 207 // IsAfter is set, look for an offset before the object, otherwise look for an 208 // offset after the object. 209 uint64_t 210 wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets, 211 bool IsAfter, uint64_t Size) { 212 // Find a minimum offset taking into account only vtable sizes. 213 uint64_t MinByte = 0; 214 for (const VirtualCallTarget &Target : Targets) { 215 if (IsAfter) 216 MinByte = std::max(MinByte, Target.minAfterBytes()); 217 else 218 MinByte = std::max(MinByte, Target.minBeforeBytes()); 219 } 220 221 // Build a vector of arrays of bytes covering, for each target, a slice of the 222 // used region (see AccumBitVector::BytesUsed in 223 // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively, 224 // this aligns the used regions to start at MinByte. 225 // 226 // In this example, A, B and C are vtables, # is a byte already allocated for 227 // a virtual function pointer, AAAA... (etc.) are the used regions for the 228 // vtables and Offset(X) is the value computed for the Offset variable below 229 // for X. 230 // 231 // Offset(A) 232 // | | 233 // |MinByte 234 // A: ################AAAAAAAA|AAAAAAAA 235 // B: ########BBBBBBBBBBBBBBBB|BBBB 236 // C: ########################|CCCCCCCCCCCCCCCC 237 // | Offset(B) | 238 // 239 // This code produces the slices of A, B and C that appear after the divider 240 // at MinByte. 241 std::vector<ArrayRef<uint8_t>> Used; 242 for (const VirtualCallTarget &Target : Targets) { 243 ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed 244 : Target.TM->Bits->Before.BytesUsed; 245 uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes() 246 : MinByte - Target.minBeforeBytes(); 247 248 // Disregard used regions that are smaller than Offset. These are 249 // effectively all-free regions that do not need to be checked. 250 if (VTUsed.size() > Offset) 251 Used.push_back(VTUsed.slice(Offset)); 252 } 253 254 if (Size == 1) { 255 // Find a free bit in each member of Used. 256 for (unsigned I = 0;; ++I) { 257 uint8_t BitsUsed = 0; 258 for (auto &&B : Used) 259 if (I < B.size()) 260 BitsUsed |= B[I]; 261 if (BitsUsed != 0xff) 262 return (MinByte + I) * 8 + 263 countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined); 264 } 265 } else { 266 // Find a free (Size/8) byte region in each member of Used. 267 // FIXME: see if alignment helps. 268 for (unsigned I = 0;; ++I) { 269 for (auto &&B : Used) { 270 unsigned Byte = 0; 271 while ((I + Byte) < B.size() && Byte < (Size / 8)) { 272 if (B[I + Byte]) 273 goto NextI; 274 ++Byte; 275 } 276 } 277 return (MinByte + I) * 8; 278 NextI:; 279 } 280 } 281 } 282 283 void wholeprogramdevirt::setBeforeReturnValues( 284 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore, 285 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 286 if (BitWidth == 1) 287 OffsetByte = -(AllocBefore / 8 + 1); 288 else 289 OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8); 290 OffsetBit = AllocBefore % 8; 291 292 for (VirtualCallTarget &Target : Targets) { 293 if (BitWidth == 1) 294 Target.setBeforeBit(AllocBefore); 295 else 296 Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8); 297 } 298 } 299 300 void wholeprogramdevirt::setAfterReturnValues( 301 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter, 302 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 303 if (BitWidth == 1) 304 OffsetByte = AllocAfter / 8; 305 else 306 OffsetByte = (AllocAfter + 7) / 8; 307 OffsetBit = AllocAfter % 8; 308 309 for (VirtualCallTarget &Target : Targets) { 310 if (BitWidth == 1) 311 Target.setAfterBit(AllocAfter); 312 else 313 Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8); 314 } 315 } 316 317 VirtualCallTarget::VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM) 318 : Fn(Fn), TM(TM), 319 IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {} 320 321 namespace { 322 323 // A slot in a set of virtual tables. The TypeID identifies the set of virtual 324 // tables, and the ByteOffset is the offset in bytes from the address point to 325 // the virtual function pointer. 326 struct VTableSlot { 327 Metadata *TypeID; 328 uint64_t ByteOffset; 329 }; 330 331 } // end anonymous namespace 332 333 namespace llvm { 334 335 template <> struct DenseMapInfo<VTableSlot> { 336 static VTableSlot getEmptyKey() { 337 return {DenseMapInfo<Metadata *>::getEmptyKey(), 338 DenseMapInfo<uint64_t>::getEmptyKey()}; 339 } 340 static VTableSlot getTombstoneKey() { 341 return {DenseMapInfo<Metadata *>::getTombstoneKey(), 342 DenseMapInfo<uint64_t>::getTombstoneKey()}; 343 } 344 static unsigned getHashValue(const VTableSlot &I) { 345 return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^ 346 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 347 } 348 static bool isEqual(const VTableSlot &LHS, 349 const VTableSlot &RHS) { 350 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 351 } 352 }; 353 354 template <> struct DenseMapInfo<VTableSlotSummary> { 355 static VTableSlotSummary getEmptyKey() { 356 return {DenseMapInfo<StringRef>::getEmptyKey(), 357 DenseMapInfo<uint64_t>::getEmptyKey()}; 358 } 359 static VTableSlotSummary getTombstoneKey() { 360 return {DenseMapInfo<StringRef>::getTombstoneKey(), 361 DenseMapInfo<uint64_t>::getTombstoneKey()}; 362 } 363 static unsigned getHashValue(const VTableSlotSummary &I) { 364 return DenseMapInfo<StringRef>::getHashValue(I.TypeID) ^ 365 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 366 } 367 static bool isEqual(const VTableSlotSummary &LHS, 368 const VTableSlotSummary &RHS) { 369 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 370 } 371 }; 372 373 } // end namespace llvm 374 375 namespace { 376 377 // Returns true if the function must be unreachable based on ValueInfo. 378 // 379 // In particular, identifies a function as unreachable in the following 380 // conditions 381 // 1) All summaries are live. 382 // 2) All function summaries indicate it's unreachable 383 bool mustBeUnreachableFunction(ValueInfo TheFnVI) { 384 if ((!TheFnVI) || TheFnVI.getSummaryList().empty()) { 385 // Returns false if ValueInfo is absent, or the summary list is empty 386 // (e.g., function declarations). 387 return false; 388 } 389 390 for (auto &Summary : TheFnVI.getSummaryList()) { 391 // Conservatively returns false if any non-live functions are seen. 392 // In general either all summaries should be live or all should be dead. 393 if (!Summary->isLive()) 394 return false; 395 if (auto *FS = dyn_cast<FunctionSummary>(Summary.get())) { 396 if (!FS->fflags().MustBeUnreachable) 397 return false; 398 } 399 // Do nothing if a non-function has the same GUID (which is rare). 400 // This is correct since non-function summaries are not relevant. 401 } 402 // All function summaries are live and all of them agree that the function is 403 // unreachble. 404 return true; 405 } 406 407 // A virtual call site. VTable is the loaded virtual table pointer, and CS is 408 // the indirect virtual call. 409 struct VirtualCallSite { 410 Value *VTable = nullptr; 411 CallBase &CB; 412 413 // If non-null, this field points to the associated unsafe use count stored in 414 // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description 415 // of that field for details. 416 unsigned *NumUnsafeUses = nullptr; 417 418 void 419 emitRemark(const StringRef OptName, const StringRef TargetName, 420 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) { 421 Function *F = CB.getCaller(); 422 DebugLoc DLoc = CB.getDebugLoc(); 423 BasicBlock *Block = CB.getParent(); 424 425 using namespace ore; 426 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, OptName, DLoc, Block) 427 << NV("Optimization", OptName) 428 << ": devirtualized a call to " 429 << NV("FunctionName", TargetName)); 430 } 431 432 void replaceAndErase( 433 const StringRef OptName, const StringRef TargetName, bool RemarksEnabled, 434 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 435 Value *New) { 436 if (RemarksEnabled) 437 emitRemark(OptName, TargetName, OREGetter); 438 CB.replaceAllUsesWith(New); 439 if (auto *II = dyn_cast<InvokeInst>(&CB)) { 440 BranchInst::Create(II->getNormalDest(), &CB); 441 II->getUnwindDest()->removePredecessor(II->getParent()); 442 } 443 CB.eraseFromParent(); 444 // This use is no longer unsafe. 445 if (NumUnsafeUses) 446 --*NumUnsafeUses; 447 } 448 }; 449 450 // Call site information collected for a specific VTableSlot and possibly a list 451 // of constant integer arguments. The grouping by arguments is handled by the 452 // VTableSlotInfo class. 453 struct CallSiteInfo { 454 /// The set of call sites for this slot. Used during regular LTO and the 455 /// import phase of ThinLTO (as well as the export phase of ThinLTO for any 456 /// call sites that appear in the merged module itself); in each of these 457 /// cases we are directly operating on the call sites at the IR level. 458 std::vector<VirtualCallSite> CallSites; 459 460 /// Whether all call sites represented by this CallSiteInfo, including those 461 /// in summaries, have been devirtualized. This starts off as true because a 462 /// default constructed CallSiteInfo represents no call sites. 463 bool AllCallSitesDevirted = true; 464 465 // These fields are used during the export phase of ThinLTO and reflect 466 // information collected from function summaries. 467 468 /// Whether any function summary contains an llvm.assume(llvm.type.test) for 469 /// this slot. 470 bool SummaryHasTypeTestAssumeUsers = false; 471 472 /// CFI-specific: a vector containing the list of function summaries that use 473 /// the llvm.type.checked.load intrinsic and therefore will require 474 /// resolutions for llvm.type.test in order to implement CFI checks if 475 /// devirtualization was unsuccessful. If devirtualization was successful, the 476 /// pass will clear this vector by calling markDevirt(). If at the end of the 477 /// pass the vector is non-empty, we will need to add a use of llvm.type.test 478 /// to each of the function summaries in the vector. 479 std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers; 480 std::vector<FunctionSummary *> SummaryTypeTestAssumeUsers; 481 482 bool isExported() const { 483 return SummaryHasTypeTestAssumeUsers || 484 !SummaryTypeCheckedLoadUsers.empty(); 485 } 486 487 void addSummaryTypeCheckedLoadUser(FunctionSummary *FS) { 488 SummaryTypeCheckedLoadUsers.push_back(FS); 489 AllCallSitesDevirted = false; 490 } 491 492 void addSummaryTypeTestAssumeUser(FunctionSummary *FS) { 493 SummaryTypeTestAssumeUsers.push_back(FS); 494 SummaryHasTypeTestAssumeUsers = true; 495 AllCallSitesDevirted = false; 496 } 497 498 void markDevirt() { 499 AllCallSitesDevirted = true; 500 501 // As explained in the comment for SummaryTypeCheckedLoadUsers. 502 SummaryTypeCheckedLoadUsers.clear(); 503 } 504 }; 505 506 // Call site information collected for a specific VTableSlot. 507 struct VTableSlotInfo { 508 // The set of call sites which do not have all constant integer arguments 509 // (excluding "this"). 510 CallSiteInfo CSInfo; 511 512 // The set of call sites with all constant integer arguments (excluding 513 // "this"), grouped by argument list. 514 std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo; 515 516 void addCallSite(Value *VTable, CallBase &CB, unsigned *NumUnsafeUses); 517 518 private: 519 CallSiteInfo &findCallSiteInfo(CallBase &CB); 520 }; 521 522 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallBase &CB) { 523 std::vector<uint64_t> Args; 524 auto *CBType = dyn_cast<IntegerType>(CB.getType()); 525 if (!CBType || CBType->getBitWidth() > 64 || CB.arg_empty()) 526 return CSInfo; 527 for (auto &&Arg : drop_begin(CB.args())) { 528 auto *CI = dyn_cast<ConstantInt>(Arg); 529 if (!CI || CI->getBitWidth() > 64) 530 return CSInfo; 531 Args.push_back(CI->getZExtValue()); 532 } 533 return ConstCSInfo[Args]; 534 } 535 536 void VTableSlotInfo::addCallSite(Value *VTable, CallBase &CB, 537 unsigned *NumUnsafeUses) { 538 auto &CSI = findCallSiteInfo(CB); 539 CSI.AllCallSitesDevirted = false; 540 CSI.CallSites.push_back({VTable, CB, NumUnsafeUses}); 541 } 542 543 struct DevirtModule { 544 Module &M; 545 function_ref<AAResults &(Function &)> AARGetter; 546 function_ref<DominatorTree &(Function &)> LookupDomTree; 547 548 ModuleSummaryIndex *ExportSummary; 549 const ModuleSummaryIndex *ImportSummary; 550 551 IntegerType *Int8Ty; 552 PointerType *Int8PtrTy; 553 IntegerType *Int32Ty; 554 IntegerType *Int64Ty; 555 IntegerType *IntPtrTy; 556 /// Sizeless array type, used for imported vtables. This provides a signal 557 /// to analyzers that these imports may alias, as they do for example 558 /// when multiple unique return values occur in the same vtable. 559 ArrayType *Int8Arr0Ty; 560 561 bool RemarksEnabled; 562 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter; 563 564 MapVector<VTableSlot, VTableSlotInfo> CallSlots; 565 566 // Calls that have already been optimized. We may add a call to multiple 567 // VTableSlotInfos if vtable loads are coalesced and need to make sure not to 568 // optimize a call more than once. 569 SmallPtrSet<CallBase *, 8> OptimizedCalls; 570 571 // This map keeps track of the number of "unsafe" uses of a loaded function 572 // pointer. The key is the associated llvm.type.test intrinsic call generated 573 // by this pass. An unsafe use is one that calls the loaded function pointer 574 // directly. Every time we eliminate an unsafe use (for example, by 575 // devirtualizing it or by applying virtual constant propagation), we 576 // decrement the value stored in this map. If a value reaches zero, we can 577 // eliminate the type check by RAUWing the associated llvm.type.test call with 578 // true. 579 std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest; 580 PatternList FunctionsToSkip; 581 582 DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter, 583 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 584 function_ref<DominatorTree &(Function &)> LookupDomTree, 585 ModuleSummaryIndex *ExportSummary, 586 const ModuleSummaryIndex *ImportSummary) 587 : M(M), AARGetter(AARGetter), LookupDomTree(LookupDomTree), 588 ExportSummary(ExportSummary), ImportSummary(ImportSummary), 589 Int8Ty(Type::getInt8Ty(M.getContext())), 590 Int8PtrTy(Type::getInt8PtrTy(M.getContext())), 591 Int32Ty(Type::getInt32Ty(M.getContext())), 592 Int64Ty(Type::getInt64Ty(M.getContext())), 593 IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)), 594 Int8Arr0Ty(ArrayType::get(Type::getInt8Ty(M.getContext()), 0)), 595 RemarksEnabled(areRemarksEnabled()), OREGetter(OREGetter) { 596 assert(!(ExportSummary && ImportSummary)); 597 FunctionsToSkip.init(SkipFunctionNames); 598 } 599 600 bool areRemarksEnabled(); 601 602 void 603 scanTypeTestUsers(Function *TypeTestFunc, 604 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap); 605 void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc); 606 607 void buildTypeIdentifierMap( 608 std::vector<VTableBits> &Bits, 609 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap); 610 611 bool 612 tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot, 613 const std::set<TypeMemberInfo> &TypeMemberInfos, 614 uint64_t ByteOffset, 615 ModuleSummaryIndex *ExportSummary); 616 617 void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn, 618 bool &IsExported); 619 bool trySingleImplDevirt(ModuleSummaryIndex *ExportSummary, 620 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 621 VTableSlotInfo &SlotInfo, 622 WholeProgramDevirtResolution *Res); 623 624 void applyICallBranchFunnel(VTableSlotInfo &SlotInfo, Constant *JT, 625 bool &IsExported); 626 void tryICallBranchFunnel(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 627 VTableSlotInfo &SlotInfo, 628 WholeProgramDevirtResolution *Res, VTableSlot Slot); 629 630 bool tryEvaluateFunctionsWithArgs( 631 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 632 ArrayRef<uint64_t> Args); 633 634 void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 635 uint64_t TheRetVal); 636 bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 637 CallSiteInfo &CSInfo, 638 WholeProgramDevirtResolution::ByArg *Res); 639 640 // Returns the global symbol name that is used to export information about the 641 // given vtable slot and list of arguments. 642 std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args, 643 StringRef Name); 644 645 bool shouldExportConstantsAsAbsoluteSymbols(); 646 647 // This function is called during the export phase to create a symbol 648 // definition containing information about the given vtable slot and list of 649 // arguments. 650 void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 651 Constant *C); 652 void exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 653 uint32_t Const, uint32_t &Storage); 654 655 // This function is called during the import phase to create a reference to 656 // the symbol definition created during the export phase. 657 Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 658 StringRef Name); 659 Constant *importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 660 StringRef Name, IntegerType *IntTy, 661 uint32_t Storage); 662 663 Constant *getMemberAddr(const TypeMemberInfo *M); 664 665 void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne, 666 Constant *UniqueMemberAddr); 667 bool tryUniqueRetValOpt(unsigned BitWidth, 668 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 669 CallSiteInfo &CSInfo, 670 WholeProgramDevirtResolution::ByArg *Res, 671 VTableSlot Slot, ArrayRef<uint64_t> Args); 672 673 void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 674 Constant *Byte, Constant *Bit); 675 bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 676 VTableSlotInfo &SlotInfo, 677 WholeProgramDevirtResolution *Res, VTableSlot Slot); 678 679 void rebuildGlobal(VTableBits &B); 680 681 // Apply the summary resolution for Slot to all virtual calls in SlotInfo. 682 void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo); 683 684 // If we were able to eliminate all unsafe uses for a type checked load, 685 // eliminate the associated type tests by replacing them with true. 686 void removeRedundantTypeTests(); 687 688 bool run(); 689 690 // Look up the corresponding ValueInfo entry of `TheFn` in `ExportSummary`. 691 // 692 // Caller guarantees that `ExportSummary` is not nullptr. 693 static ValueInfo lookUpFunctionValueInfo(Function *TheFn, 694 ModuleSummaryIndex *ExportSummary); 695 696 // Returns true if the function definition must be unreachable. 697 // 698 // Note if this helper function returns true, `F` is guaranteed 699 // to be unreachable; if it returns false, `F` might still 700 // be unreachable but not covered by this helper function. 701 // 702 // Implementation-wise, if function definition is present, IR is analyzed; if 703 // not, look up function flags from ExportSummary as a fallback. 704 static bool mustBeUnreachableFunction(Function *const F, 705 ModuleSummaryIndex *ExportSummary); 706 707 // Lower the module using the action and summary passed as command line 708 // arguments. For testing purposes only. 709 static bool 710 runForTesting(Module &M, function_ref<AAResults &(Function &)> AARGetter, 711 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 712 function_ref<DominatorTree &(Function &)> LookupDomTree); 713 }; 714 715 struct DevirtIndex { 716 ModuleSummaryIndex &ExportSummary; 717 // The set in which to record GUIDs exported from their module by 718 // devirtualization, used by client to ensure they are not internalized. 719 std::set<GlobalValue::GUID> &ExportedGUIDs; 720 // A map in which to record the information necessary to locate the WPD 721 // resolution for local targets in case they are exported by cross module 722 // importing. 723 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap; 724 725 MapVector<VTableSlotSummary, VTableSlotInfo> CallSlots; 726 727 PatternList FunctionsToSkip; 728 729 DevirtIndex( 730 ModuleSummaryIndex &ExportSummary, 731 std::set<GlobalValue::GUID> &ExportedGUIDs, 732 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) 733 : ExportSummary(ExportSummary), ExportedGUIDs(ExportedGUIDs), 734 LocalWPDTargetsMap(LocalWPDTargetsMap) { 735 FunctionsToSkip.init(SkipFunctionNames); 736 } 737 738 bool tryFindVirtualCallTargets(std::vector<ValueInfo> &TargetsForSlot, 739 const TypeIdCompatibleVtableInfo TIdInfo, 740 uint64_t ByteOffset); 741 742 bool trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 743 VTableSlotSummary &SlotSummary, 744 VTableSlotInfo &SlotInfo, 745 WholeProgramDevirtResolution *Res, 746 std::set<ValueInfo> &DevirtTargets); 747 748 void run(); 749 }; 750 751 struct WholeProgramDevirt : public ModulePass { 752 static char ID; 753 754 bool UseCommandLine = false; 755 756 ModuleSummaryIndex *ExportSummary = nullptr; 757 const ModuleSummaryIndex *ImportSummary = nullptr; 758 759 WholeProgramDevirt() : ModulePass(ID), UseCommandLine(true) { 760 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry()); 761 } 762 763 WholeProgramDevirt(ModuleSummaryIndex *ExportSummary, 764 const ModuleSummaryIndex *ImportSummary) 765 : ModulePass(ID), ExportSummary(ExportSummary), 766 ImportSummary(ImportSummary) { 767 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry()); 768 } 769 770 bool runOnModule(Module &M) override { 771 if (skipModule(M)) 772 return false; 773 774 // In the new pass manager, we can request the optimization 775 // remark emitter pass on a per-function-basis, which the 776 // OREGetter will do for us. 777 // In the old pass manager, this is harder, so we just build 778 // an optimization remark emitter on the fly, when we need it. 779 std::unique_ptr<OptimizationRemarkEmitter> ORE; 780 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { 781 ORE = std::make_unique<OptimizationRemarkEmitter>(F); 782 return *ORE; 783 }; 784 785 auto LookupDomTree = [this](Function &F) -> DominatorTree & { 786 return this->getAnalysis<DominatorTreeWrapperPass>(F).getDomTree(); 787 }; 788 789 if (UseCommandLine) 790 return DevirtModule::runForTesting(M, LegacyAARGetter(*this), OREGetter, 791 LookupDomTree); 792 793 return DevirtModule(M, LegacyAARGetter(*this), OREGetter, LookupDomTree, 794 ExportSummary, ImportSummary) 795 .run(); 796 } 797 798 void getAnalysisUsage(AnalysisUsage &AU) const override { 799 AU.addRequired<AssumptionCacheTracker>(); 800 AU.addRequired<TargetLibraryInfoWrapperPass>(); 801 AU.addRequired<DominatorTreeWrapperPass>(); 802 } 803 }; 804 805 } // end anonymous namespace 806 807 INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt", 808 "Whole program devirtualization", false, false) 809 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 810 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 811 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 812 INITIALIZE_PASS_END(WholeProgramDevirt, "wholeprogramdevirt", 813 "Whole program devirtualization", false, false) 814 char WholeProgramDevirt::ID = 0; 815 816 ModulePass * 817 llvm::createWholeProgramDevirtPass(ModuleSummaryIndex *ExportSummary, 818 const ModuleSummaryIndex *ImportSummary) { 819 return new WholeProgramDevirt(ExportSummary, ImportSummary); 820 } 821 822 PreservedAnalyses WholeProgramDevirtPass::run(Module &M, 823 ModuleAnalysisManager &AM) { 824 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 825 auto AARGetter = [&](Function &F) -> AAResults & { 826 return FAM.getResult<AAManager>(F); 827 }; 828 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { 829 return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F); 830 }; 831 auto LookupDomTree = [&FAM](Function &F) -> DominatorTree & { 832 return FAM.getResult<DominatorTreeAnalysis>(F); 833 }; 834 if (UseCommandLine) { 835 if (DevirtModule::runForTesting(M, AARGetter, OREGetter, LookupDomTree)) 836 return PreservedAnalyses::all(); 837 return PreservedAnalyses::none(); 838 } 839 if (!DevirtModule(M, AARGetter, OREGetter, LookupDomTree, ExportSummary, 840 ImportSummary) 841 .run()) 842 return PreservedAnalyses::all(); 843 return PreservedAnalyses::none(); 844 } 845 846 // Enable whole program visibility if enabled by client (e.g. linker) or 847 // internal option, and not force disabled. 848 static bool hasWholeProgramVisibility(bool WholeProgramVisibilityEnabledInLTO) { 849 return (WholeProgramVisibilityEnabledInLTO || WholeProgramVisibility) && 850 !DisableWholeProgramVisibility; 851 } 852 853 namespace llvm { 854 855 /// If whole program visibility asserted, then upgrade all public vcall 856 /// visibility metadata on vtable definitions to linkage unit visibility in 857 /// Module IR (for regular or hybrid LTO). 858 void updateVCallVisibilityInModule( 859 Module &M, bool WholeProgramVisibilityEnabledInLTO, 860 const DenseSet<GlobalValue::GUID> &DynamicExportSymbols) { 861 if (!hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO)) 862 return; 863 for (GlobalVariable &GV : M.globals()) 864 // Add linkage unit visibility to any variable with type metadata, which are 865 // the vtable definitions. We won't have an existing vcall_visibility 866 // metadata on vtable definitions with public visibility. 867 if (GV.hasMetadata(LLVMContext::MD_type) && 868 GV.getVCallVisibility() == GlobalObject::VCallVisibilityPublic && 869 // Don't upgrade the visibility for symbols exported to the dynamic 870 // linker, as we have no information on their eventual use. 871 !DynamicExportSymbols.count(GV.getGUID())) 872 GV.setVCallVisibilityMetadata(GlobalObject::VCallVisibilityLinkageUnit); 873 } 874 875 /// If whole program visibility asserted, then upgrade all public vcall 876 /// visibility metadata on vtable definition summaries to linkage unit 877 /// visibility in Module summary index (for ThinLTO). 878 void updateVCallVisibilityInIndex( 879 ModuleSummaryIndex &Index, bool WholeProgramVisibilityEnabledInLTO, 880 const DenseSet<GlobalValue::GUID> &DynamicExportSymbols) { 881 if (!hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO)) 882 return; 883 for (auto &P : Index) { 884 // Don't upgrade the visibility for symbols exported to the dynamic 885 // linker, as we have no information on their eventual use. 886 if (DynamicExportSymbols.count(P.first)) 887 continue; 888 for (auto &S : P.second.SummaryList) { 889 auto *GVar = dyn_cast<GlobalVarSummary>(S.get()); 890 if (!GVar || 891 GVar->getVCallVisibility() != GlobalObject::VCallVisibilityPublic) 892 continue; 893 GVar->setVCallVisibility(GlobalObject::VCallVisibilityLinkageUnit); 894 } 895 } 896 } 897 898 void runWholeProgramDevirtOnIndex( 899 ModuleSummaryIndex &Summary, std::set<GlobalValue::GUID> &ExportedGUIDs, 900 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 901 DevirtIndex(Summary, ExportedGUIDs, LocalWPDTargetsMap).run(); 902 } 903 904 void updateIndexWPDForExports( 905 ModuleSummaryIndex &Summary, 906 function_ref<bool(StringRef, ValueInfo)> isExported, 907 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 908 for (auto &T : LocalWPDTargetsMap) { 909 auto &VI = T.first; 910 // This was enforced earlier during trySingleImplDevirt. 911 assert(VI.getSummaryList().size() == 1 && 912 "Devirt of local target has more than one copy"); 913 auto &S = VI.getSummaryList()[0]; 914 if (!isExported(S->modulePath(), VI)) 915 continue; 916 917 // It's been exported by a cross module import. 918 for (auto &SlotSummary : T.second) { 919 auto *TIdSum = Summary.getTypeIdSummary(SlotSummary.TypeID); 920 assert(TIdSum); 921 auto WPDRes = TIdSum->WPDRes.find(SlotSummary.ByteOffset); 922 assert(WPDRes != TIdSum->WPDRes.end()); 923 WPDRes->second.SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 924 WPDRes->second.SingleImplName, 925 Summary.getModuleHash(S->modulePath())); 926 } 927 } 928 } 929 930 } // end namespace llvm 931 932 static Error checkCombinedSummaryForTesting(ModuleSummaryIndex *Summary) { 933 // Check that summary index contains regular LTO module when performing 934 // export to prevent occasional use of index from pure ThinLTO compilation 935 // (-fno-split-lto-module). This kind of summary index is passed to 936 // DevirtIndex::run, not to DevirtModule::run used by opt/runForTesting. 937 const auto &ModPaths = Summary->modulePaths(); 938 if (ClSummaryAction != PassSummaryAction::Import && 939 ModPaths.find(ModuleSummaryIndex::getRegularLTOModuleName()) == 940 ModPaths.end()) 941 return createStringError( 942 errc::invalid_argument, 943 "combined summary should contain Regular LTO module"); 944 return ErrorSuccess(); 945 } 946 947 bool DevirtModule::runForTesting( 948 Module &M, function_ref<AAResults &(Function &)> AARGetter, 949 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 950 function_ref<DominatorTree &(Function &)> LookupDomTree) { 951 std::unique_ptr<ModuleSummaryIndex> Summary = 952 std::make_unique<ModuleSummaryIndex>(/*HaveGVs=*/false); 953 954 // Handle the command-line summary arguments. This code is for testing 955 // purposes only, so we handle errors directly. 956 if (!ClReadSummary.empty()) { 957 ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary + 958 ": "); 959 auto ReadSummaryFile = 960 ExitOnErr(errorOrToExpected(MemoryBuffer::getFile(ClReadSummary))); 961 if (Expected<std::unique_ptr<ModuleSummaryIndex>> SummaryOrErr = 962 getModuleSummaryIndex(*ReadSummaryFile)) { 963 Summary = std::move(*SummaryOrErr); 964 ExitOnErr(checkCombinedSummaryForTesting(Summary.get())); 965 } else { 966 // Try YAML if we've failed with bitcode. 967 consumeError(SummaryOrErr.takeError()); 968 yaml::Input In(ReadSummaryFile->getBuffer()); 969 In >> *Summary; 970 ExitOnErr(errorCodeToError(In.error())); 971 } 972 } 973 974 bool Changed = 975 DevirtModule(M, AARGetter, OREGetter, LookupDomTree, 976 ClSummaryAction == PassSummaryAction::Export ? Summary.get() 977 : nullptr, 978 ClSummaryAction == PassSummaryAction::Import ? Summary.get() 979 : nullptr) 980 .run(); 981 982 if (!ClWriteSummary.empty()) { 983 ExitOnError ExitOnErr( 984 "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": "); 985 std::error_code EC; 986 if (StringRef(ClWriteSummary).endswith(".bc")) { 987 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_None); 988 ExitOnErr(errorCodeToError(EC)); 989 writeIndexToFile(*Summary, OS); 990 } else { 991 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_TextWithCRLF); 992 ExitOnErr(errorCodeToError(EC)); 993 yaml::Output Out(OS); 994 Out << *Summary; 995 } 996 } 997 998 return Changed; 999 } 1000 1001 void DevirtModule::buildTypeIdentifierMap( 1002 std::vector<VTableBits> &Bits, 1003 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 1004 DenseMap<GlobalVariable *, VTableBits *> GVToBits; 1005 Bits.reserve(M.getGlobalList().size()); 1006 SmallVector<MDNode *, 2> Types; 1007 for (GlobalVariable &GV : M.globals()) { 1008 Types.clear(); 1009 GV.getMetadata(LLVMContext::MD_type, Types); 1010 if (GV.isDeclaration() || Types.empty()) 1011 continue; 1012 1013 VTableBits *&BitsPtr = GVToBits[&GV]; 1014 if (!BitsPtr) { 1015 Bits.emplace_back(); 1016 Bits.back().GV = &GV; 1017 Bits.back().ObjectSize = 1018 M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType()); 1019 BitsPtr = &Bits.back(); 1020 } 1021 1022 for (MDNode *Type : Types) { 1023 auto TypeID = Type->getOperand(1).get(); 1024 1025 uint64_t Offset = 1026 cast<ConstantInt>( 1027 cast<ConstantAsMetadata>(Type->getOperand(0))->getValue()) 1028 ->getZExtValue(); 1029 1030 TypeIdMap[TypeID].insert({BitsPtr, Offset}); 1031 } 1032 } 1033 } 1034 1035 bool DevirtModule::tryFindVirtualCallTargets( 1036 std::vector<VirtualCallTarget> &TargetsForSlot, 1037 const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset, 1038 ModuleSummaryIndex *ExportSummary) { 1039 for (const TypeMemberInfo &TM : TypeMemberInfos) { 1040 if (!TM.Bits->GV->isConstant()) 1041 return false; 1042 1043 // We cannot perform whole program devirtualization analysis on a vtable 1044 // with public LTO visibility. 1045 if (TM.Bits->GV->getVCallVisibility() == 1046 GlobalObject::VCallVisibilityPublic) 1047 return false; 1048 1049 Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(), 1050 TM.Offset + ByteOffset, M); 1051 if (!Ptr) 1052 return false; 1053 1054 auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts()); 1055 if (!Fn) 1056 return false; 1057 1058 if (FunctionsToSkip.match(Fn->getName())) 1059 return false; 1060 1061 // We can disregard __cxa_pure_virtual as a possible call target, as 1062 // calls to pure virtuals are UB. 1063 if (Fn->getName() == "__cxa_pure_virtual") 1064 continue; 1065 1066 // We can disregard unreachable functions as possible call targets, as 1067 // unreachable functions shouldn't be called. 1068 if (mustBeUnreachableFunction(Fn, ExportSummary)) 1069 continue; 1070 1071 TargetsForSlot.push_back({Fn, &TM}); 1072 } 1073 1074 // Give up if we couldn't find any targets. 1075 return !TargetsForSlot.empty(); 1076 } 1077 1078 bool DevirtIndex::tryFindVirtualCallTargets( 1079 std::vector<ValueInfo> &TargetsForSlot, const TypeIdCompatibleVtableInfo TIdInfo, 1080 uint64_t ByteOffset) { 1081 for (const TypeIdOffsetVtableInfo &P : TIdInfo) { 1082 // Find a representative copy of the vtable initializer. 1083 // We can have multiple available_externally, linkonce_odr and weak_odr 1084 // vtable initializers. We can also have multiple external vtable 1085 // initializers in the case of comdats, which we cannot check here. 1086 // The linker should give an error in this case. 1087 // 1088 // Also, handle the case of same-named local Vtables with the same path 1089 // and therefore the same GUID. This can happen if there isn't enough 1090 // distinguishing path when compiling the source file. In that case we 1091 // conservatively return false early. 1092 const GlobalVarSummary *VS = nullptr; 1093 bool LocalFound = false; 1094 for (auto &S : P.VTableVI.getSummaryList()) { 1095 if (GlobalValue::isLocalLinkage(S->linkage())) { 1096 if (LocalFound) 1097 return false; 1098 LocalFound = true; 1099 } 1100 auto *CurVS = cast<GlobalVarSummary>(S->getBaseObject()); 1101 if (!CurVS->vTableFuncs().empty() || 1102 // Previously clang did not attach the necessary type metadata to 1103 // available_externally vtables, in which case there would not 1104 // be any vtable functions listed in the summary and we need 1105 // to treat this case conservatively (in case the bitcode is old). 1106 // However, we will also not have any vtable functions in the 1107 // case of a pure virtual base class. In that case we do want 1108 // to set VS to avoid treating it conservatively. 1109 !GlobalValue::isAvailableExternallyLinkage(S->linkage())) { 1110 VS = CurVS; 1111 // We cannot perform whole program devirtualization analysis on a vtable 1112 // with public LTO visibility. 1113 if (VS->getVCallVisibility() == GlobalObject::VCallVisibilityPublic) 1114 return false; 1115 } 1116 } 1117 // There will be no VS if all copies are available_externally having no 1118 // type metadata. In that case we can't safely perform WPD. 1119 if (!VS) 1120 return false; 1121 if (!VS->isLive()) 1122 continue; 1123 for (auto VTP : VS->vTableFuncs()) { 1124 if (VTP.VTableOffset != P.AddressPointOffset + ByteOffset) 1125 continue; 1126 1127 if (mustBeUnreachableFunction(VTP.FuncVI)) 1128 continue; 1129 1130 TargetsForSlot.push_back(VTP.FuncVI); 1131 } 1132 } 1133 1134 // Give up if we couldn't find any targets. 1135 return !TargetsForSlot.empty(); 1136 } 1137 1138 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo, 1139 Constant *TheFn, bool &IsExported) { 1140 // Don't devirtualize function if we're told to skip it 1141 // in -wholeprogramdevirt-skip. 1142 if (FunctionsToSkip.match(TheFn->stripPointerCasts()->getName())) 1143 return; 1144 auto Apply = [&](CallSiteInfo &CSInfo) { 1145 for (auto &&VCallSite : CSInfo.CallSites) { 1146 if (!OptimizedCalls.insert(&VCallSite.CB).second) 1147 continue; 1148 1149 if (RemarksEnabled) 1150 VCallSite.emitRemark("single-impl", 1151 TheFn->stripPointerCasts()->getName(), OREGetter); 1152 NumSingleImpl++; 1153 auto &CB = VCallSite.CB; 1154 assert(!CB.getCalledFunction() && "devirtualizing direct call?"); 1155 IRBuilder<> Builder(&CB); 1156 Value *Callee = 1157 Builder.CreateBitCast(TheFn, CB.getCalledOperand()->getType()); 1158 1159 // If trap checking is enabled, add support to compare the virtual 1160 // function pointer to the devirtualized target. In case of a mismatch, 1161 // perform a debug trap. 1162 if (DevirtCheckMode == WPDCheckMode::Trap) { 1163 auto *Cond = Builder.CreateICmpNE(CB.getCalledOperand(), Callee); 1164 Instruction *ThenTerm = 1165 SplitBlockAndInsertIfThen(Cond, &CB, /*Unreachable=*/false); 1166 Builder.SetInsertPoint(ThenTerm); 1167 Function *TrapFn = Intrinsic::getDeclaration(&M, Intrinsic::debugtrap); 1168 auto *CallTrap = Builder.CreateCall(TrapFn); 1169 CallTrap->setDebugLoc(CB.getDebugLoc()); 1170 } 1171 1172 // If fallback checking is enabled, add support to compare the virtual 1173 // function pointer to the devirtualized target. In case of a mismatch, 1174 // fall back to indirect call. 1175 if (DevirtCheckMode == WPDCheckMode::Fallback) { 1176 MDNode *Weights = 1177 MDBuilder(M.getContext()).createBranchWeights((1U << 20) - 1, 1); 1178 // Version the indirect call site. If the called value is equal to the 1179 // given callee, 'NewInst' will be executed, otherwise the original call 1180 // site will be executed. 1181 CallBase &NewInst = versionCallSite(CB, Callee, Weights); 1182 NewInst.setCalledOperand(Callee); 1183 // Since the new call site is direct, we must clear metadata that 1184 // is only appropriate for indirect calls. This includes !prof and 1185 // !callees metadata. 1186 NewInst.setMetadata(LLVMContext::MD_prof, nullptr); 1187 NewInst.setMetadata(LLVMContext::MD_callees, nullptr); 1188 // Additionally, we should remove them from the fallback indirect call, 1189 // so that we don't attempt to perform indirect call promotion later. 1190 CB.setMetadata(LLVMContext::MD_prof, nullptr); 1191 CB.setMetadata(LLVMContext::MD_callees, nullptr); 1192 } 1193 1194 // In either trapping or non-checking mode, devirtualize original call. 1195 else { 1196 // Devirtualize unconditionally. 1197 CB.setCalledOperand(Callee); 1198 // Since the call site is now direct, we must clear metadata that 1199 // is only appropriate for indirect calls. This includes !prof and 1200 // !callees metadata. 1201 CB.setMetadata(LLVMContext::MD_prof, nullptr); 1202 CB.setMetadata(LLVMContext::MD_callees, nullptr); 1203 } 1204 1205 // This use is no longer unsafe. 1206 if (VCallSite.NumUnsafeUses) 1207 --*VCallSite.NumUnsafeUses; 1208 } 1209 if (CSInfo.isExported()) 1210 IsExported = true; 1211 CSInfo.markDevirt(); 1212 }; 1213 Apply(SlotInfo.CSInfo); 1214 for (auto &P : SlotInfo.ConstCSInfo) 1215 Apply(P.second); 1216 } 1217 1218 static bool AddCalls(VTableSlotInfo &SlotInfo, const ValueInfo &Callee) { 1219 // We can't add calls if we haven't seen a definition 1220 if (Callee.getSummaryList().empty()) 1221 return false; 1222 1223 // Insert calls into the summary index so that the devirtualized targets 1224 // are eligible for import. 1225 // FIXME: Annotate type tests with hotness. For now, mark these as hot 1226 // to better ensure we have the opportunity to inline them. 1227 bool IsExported = false; 1228 auto &S = Callee.getSummaryList()[0]; 1229 CalleeInfo CI(CalleeInfo::HotnessType::Hot, /* RelBF = */ 0); 1230 auto AddCalls = [&](CallSiteInfo &CSInfo) { 1231 for (auto *FS : CSInfo.SummaryTypeCheckedLoadUsers) { 1232 FS->addCall({Callee, CI}); 1233 IsExported |= S->modulePath() != FS->modulePath(); 1234 } 1235 for (auto *FS : CSInfo.SummaryTypeTestAssumeUsers) { 1236 FS->addCall({Callee, CI}); 1237 IsExported |= S->modulePath() != FS->modulePath(); 1238 } 1239 }; 1240 AddCalls(SlotInfo.CSInfo); 1241 for (auto &P : SlotInfo.ConstCSInfo) 1242 AddCalls(P.second); 1243 return IsExported; 1244 } 1245 1246 bool DevirtModule::trySingleImplDevirt( 1247 ModuleSummaryIndex *ExportSummary, 1248 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1249 WholeProgramDevirtResolution *Res) { 1250 // See if the program contains a single implementation of this virtual 1251 // function. 1252 Function *TheFn = TargetsForSlot[0].Fn; 1253 for (auto &&Target : TargetsForSlot) 1254 if (TheFn != Target.Fn) 1255 return false; 1256 1257 // If so, update each call site to call that implementation directly. 1258 if (RemarksEnabled || AreStatisticsEnabled()) 1259 TargetsForSlot[0].WasDevirt = true; 1260 1261 bool IsExported = false; 1262 applySingleImplDevirt(SlotInfo, TheFn, IsExported); 1263 if (!IsExported) 1264 return false; 1265 1266 // If the only implementation has local linkage, we must promote to external 1267 // to make it visible to thin LTO objects. We can only get here during the 1268 // ThinLTO export phase. 1269 if (TheFn->hasLocalLinkage()) { 1270 std::string NewName = (TheFn->getName() + ".llvm.merged").str(); 1271 1272 // Since we are renaming the function, any comdats with the same name must 1273 // also be renamed. This is required when targeting COFF, as the comdat name 1274 // must match one of the names of the symbols in the comdat. 1275 if (Comdat *C = TheFn->getComdat()) { 1276 if (C->getName() == TheFn->getName()) { 1277 Comdat *NewC = M.getOrInsertComdat(NewName); 1278 NewC->setSelectionKind(C->getSelectionKind()); 1279 for (GlobalObject &GO : M.global_objects()) 1280 if (GO.getComdat() == C) 1281 GO.setComdat(NewC); 1282 } 1283 } 1284 1285 TheFn->setLinkage(GlobalValue::ExternalLinkage); 1286 TheFn->setVisibility(GlobalValue::HiddenVisibility); 1287 TheFn->setName(NewName); 1288 } 1289 if (ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFn->getGUID())) 1290 // Any needed promotion of 'TheFn' has already been done during 1291 // LTO unit split, so we can ignore return value of AddCalls. 1292 AddCalls(SlotInfo, TheFnVI); 1293 1294 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1295 Res->SingleImplName = std::string(TheFn->getName()); 1296 1297 return true; 1298 } 1299 1300 bool DevirtIndex::trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 1301 VTableSlotSummary &SlotSummary, 1302 VTableSlotInfo &SlotInfo, 1303 WholeProgramDevirtResolution *Res, 1304 std::set<ValueInfo> &DevirtTargets) { 1305 // See if the program contains a single implementation of this virtual 1306 // function. 1307 auto TheFn = TargetsForSlot[0]; 1308 for (auto &&Target : TargetsForSlot) 1309 if (TheFn != Target) 1310 return false; 1311 1312 // Don't devirtualize if we don't have target definition. 1313 auto Size = TheFn.getSummaryList().size(); 1314 if (!Size) 1315 return false; 1316 1317 // Don't devirtualize function if we're told to skip it 1318 // in -wholeprogramdevirt-skip. 1319 if (FunctionsToSkip.match(TheFn.name())) 1320 return false; 1321 1322 // If the summary list contains multiple summaries where at least one is 1323 // a local, give up, as we won't know which (possibly promoted) name to use. 1324 for (auto &S : TheFn.getSummaryList()) 1325 if (GlobalValue::isLocalLinkage(S->linkage()) && Size > 1) 1326 return false; 1327 1328 // Collect functions devirtualized at least for one call site for stats. 1329 if (PrintSummaryDevirt || AreStatisticsEnabled()) 1330 DevirtTargets.insert(TheFn); 1331 1332 auto &S = TheFn.getSummaryList()[0]; 1333 bool IsExported = AddCalls(SlotInfo, TheFn); 1334 if (IsExported) 1335 ExportedGUIDs.insert(TheFn.getGUID()); 1336 1337 // Record in summary for use in devirtualization during the ThinLTO import 1338 // step. 1339 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1340 if (GlobalValue::isLocalLinkage(S->linkage())) { 1341 if (IsExported) 1342 // If target is a local function and we are exporting it by 1343 // devirtualizing a call in another module, we need to record the 1344 // promoted name. 1345 Res->SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 1346 TheFn.name(), ExportSummary.getModuleHash(S->modulePath())); 1347 else { 1348 LocalWPDTargetsMap[TheFn].push_back(SlotSummary); 1349 Res->SingleImplName = std::string(TheFn.name()); 1350 } 1351 } else 1352 Res->SingleImplName = std::string(TheFn.name()); 1353 1354 // Name will be empty if this thin link driven off of serialized combined 1355 // index (e.g. llvm-lto). However, WPD is not supported/invoked for the 1356 // legacy LTO API anyway. 1357 assert(!Res->SingleImplName.empty()); 1358 1359 return true; 1360 } 1361 1362 void DevirtModule::tryICallBranchFunnel( 1363 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1364 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1365 Triple T(M.getTargetTriple()); 1366 if (T.getArch() != Triple::x86_64) 1367 return; 1368 1369 if (TargetsForSlot.size() > ClThreshold) 1370 return; 1371 1372 bool HasNonDevirt = !SlotInfo.CSInfo.AllCallSitesDevirted; 1373 if (!HasNonDevirt) 1374 for (auto &P : SlotInfo.ConstCSInfo) 1375 if (!P.second.AllCallSitesDevirted) { 1376 HasNonDevirt = true; 1377 break; 1378 } 1379 1380 if (!HasNonDevirt) 1381 return; 1382 1383 FunctionType *FT = 1384 FunctionType::get(Type::getVoidTy(M.getContext()), {Int8PtrTy}, true); 1385 Function *JT; 1386 if (isa<MDString>(Slot.TypeID)) { 1387 JT = Function::Create(FT, Function::ExternalLinkage, 1388 M.getDataLayout().getProgramAddressSpace(), 1389 getGlobalName(Slot, {}, "branch_funnel"), &M); 1390 JT->setVisibility(GlobalValue::HiddenVisibility); 1391 } else { 1392 JT = Function::Create(FT, Function::InternalLinkage, 1393 M.getDataLayout().getProgramAddressSpace(), 1394 "branch_funnel", &M); 1395 } 1396 JT->addParamAttr(0, Attribute::Nest); 1397 1398 std::vector<Value *> JTArgs; 1399 JTArgs.push_back(JT->arg_begin()); 1400 for (auto &T : TargetsForSlot) { 1401 JTArgs.push_back(getMemberAddr(T.TM)); 1402 JTArgs.push_back(T.Fn); 1403 } 1404 1405 BasicBlock *BB = BasicBlock::Create(M.getContext(), "", JT, nullptr); 1406 Function *Intr = 1407 Intrinsic::getDeclaration(&M, llvm::Intrinsic::icall_branch_funnel, {}); 1408 1409 auto *CI = CallInst::Create(Intr, JTArgs, "", BB); 1410 CI->setTailCallKind(CallInst::TCK_MustTail); 1411 ReturnInst::Create(M.getContext(), nullptr, BB); 1412 1413 bool IsExported = false; 1414 applyICallBranchFunnel(SlotInfo, JT, IsExported); 1415 if (IsExported) 1416 Res->TheKind = WholeProgramDevirtResolution::BranchFunnel; 1417 } 1418 1419 void DevirtModule::applyICallBranchFunnel(VTableSlotInfo &SlotInfo, 1420 Constant *JT, bool &IsExported) { 1421 auto Apply = [&](CallSiteInfo &CSInfo) { 1422 if (CSInfo.isExported()) 1423 IsExported = true; 1424 if (CSInfo.AllCallSitesDevirted) 1425 return; 1426 for (auto &&VCallSite : CSInfo.CallSites) { 1427 CallBase &CB = VCallSite.CB; 1428 1429 // Jump tables are only profitable if the retpoline mitigation is enabled. 1430 Attribute FSAttr = CB.getCaller()->getFnAttribute("target-features"); 1431 if (!FSAttr.isValid() || 1432 !FSAttr.getValueAsString().contains("+retpoline")) 1433 continue; 1434 1435 NumBranchFunnel++; 1436 if (RemarksEnabled) 1437 VCallSite.emitRemark("branch-funnel", 1438 JT->stripPointerCasts()->getName(), OREGetter); 1439 1440 // Pass the address of the vtable in the nest register, which is r10 on 1441 // x86_64. 1442 std::vector<Type *> NewArgs; 1443 NewArgs.push_back(Int8PtrTy); 1444 append_range(NewArgs, CB.getFunctionType()->params()); 1445 FunctionType *NewFT = 1446 FunctionType::get(CB.getFunctionType()->getReturnType(), NewArgs, 1447 CB.getFunctionType()->isVarArg()); 1448 PointerType *NewFTPtr = PointerType::getUnqual(NewFT); 1449 1450 IRBuilder<> IRB(&CB); 1451 std::vector<Value *> Args; 1452 Args.push_back(IRB.CreateBitCast(VCallSite.VTable, Int8PtrTy)); 1453 llvm::append_range(Args, CB.args()); 1454 1455 CallBase *NewCS = nullptr; 1456 if (isa<CallInst>(CB)) 1457 NewCS = IRB.CreateCall(NewFT, IRB.CreateBitCast(JT, NewFTPtr), Args); 1458 else 1459 NewCS = IRB.CreateInvoke(NewFT, IRB.CreateBitCast(JT, NewFTPtr), 1460 cast<InvokeInst>(CB).getNormalDest(), 1461 cast<InvokeInst>(CB).getUnwindDest(), Args); 1462 NewCS->setCallingConv(CB.getCallingConv()); 1463 1464 AttributeList Attrs = CB.getAttributes(); 1465 std::vector<AttributeSet> NewArgAttrs; 1466 NewArgAttrs.push_back(AttributeSet::get( 1467 M.getContext(), ArrayRef<Attribute>{Attribute::get( 1468 M.getContext(), Attribute::Nest)})); 1469 for (unsigned I = 0; I + 2 < Attrs.getNumAttrSets(); ++I) 1470 NewArgAttrs.push_back(Attrs.getParamAttrs(I)); 1471 NewCS->setAttributes( 1472 AttributeList::get(M.getContext(), Attrs.getFnAttrs(), 1473 Attrs.getRetAttrs(), NewArgAttrs)); 1474 1475 CB.replaceAllUsesWith(NewCS); 1476 CB.eraseFromParent(); 1477 1478 // This use is no longer unsafe. 1479 if (VCallSite.NumUnsafeUses) 1480 --*VCallSite.NumUnsafeUses; 1481 } 1482 // Don't mark as devirtualized because there may be callers compiled without 1483 // retpoline mitigation, which would mean that they are lowered to 1484 // llvm.type.test and therefore require an llvm.type.test resolution for the 1485 // type identifier. 1486 }; 1487 Apply(SlotInfo.CSInfo); 1488 for (auto &P : SlotInfo.ConstCSInfo) 1489 Apply(P.second); 1490 } 1491 1492 bool DevirtModule::tryEvaluateFunctionsWithArgs( 1493 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1494 ArrayRef<uint64_t> Args) { 1495 // Evaluate each function and store the result in each target's RetVal 1496 // field. 1497 for (VirtualCallTarget &Target : TargetsForSlot) { 1498 if (Target.Fn->arg_size() != Args.size() + 1) 1499 return false; 1500 1501 Evaluator Eval(M.getDataLayout(), nullptr); 1502 SmallVector<Constant *, 2> EvalArgs; 1503 EvalArgs.push_back( 1504 Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0))); 1505 for (unsigned I = 0; I != Args.size(); ++I) { 1506 auto *ArgTy = dyn_cast<IntegerType>( 1507 Target.Fn->getFunctionType()->getParamType(I + 1)); 1508 if (!ArgTy) 1509 return false; 1510 EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I])); 1511 } 1512 1513 Constant *RetVal; 1514 if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) || 1515 !isa<ConstantInt>(RetVal)) 1516 return false; 1517 Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue(); 1518 } 1519 return true; 1520 } 1521 1522 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1523 uint64_t TheRetVal) { 1524 for (auto Call : CSInfo.CallSites) { 1525 if (!OptimizedCalls.insert(&Call.CB).second) 1526 continue; 1527 NumUniformRetVal++; 1528 Call.replaceAndErase( 1529 "uniform-ret-val", FnName, RemarksEnabled, OREGetter, 1530 ConstantInt::get(cast<IntegerType>(Call.CB.getType()), TheRetVal)); 1531 } 1532 CSInfo.markDevirt(); 1533 } 1534 1535 bool DevirtModule::tryUniformRetValOpt( 1536 MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo, 1537 WholeProgramDevirtResolution::ByArg *Res) { 1538 // Uniform return value optimization. If all functions return the same 1539 // constant, replace all calls with that constant. 1540 uint64_t TheRetVal = TargetsForSlot[0].RetVal; 1541 for (const VirtualCallTarget &Target : TargetsForSlot) 1542 if (Target.RetVal != TheRetVal) 1543 return false; 1544 1545 if (CSInfo.isExported()) { 1546 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniformRetVal; 1547 Res->Info = TheRetVal; 1548 } 1549 1550 applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal); 1551 if (RemarksEnabled || AreStatisticsEnabled()) 1552 for (auto &&Target : TargetsForSlot) 1553 Target.WasDevirt = true; 1554 return true; 1555 } 1556 1557 std::string DevirtModule::getGlobalName(VTableSlot Slot, 1558 ArrayRef<uint64_t> Args, 1559 StringRef Name) { 1560 std::string FullName = "__typeid_"; 1561 raw_string_ostream OS(FullName); 1562 OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset; 1563 for (uint64_t Arg : Args) 1564 OS << '_' << Arg; 1565 OS << '_' << Name; 1566 return OS.str(); 1567 } 1568 1569 bool DevirtModule::shouldExportConstantsAsAbsoluteSymbols() { 1570 Triple T(M.getTargetTriple()); 1571 return T.isX86() && T.getObjectFormat() == Triple::ELF; 1572 } 1573 1574 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1575 StringRef Name, Constant *C) { 1576 GlobalAlias *GA = GlobalAlias::create(Int8Ty, 0, GlobalValue::ExternalLinkage, 1577 getGlobalName(Slot, Args, Name), C, &M); 1578 GA->setVisibility(GlobalValue::HiddenVisibility); 1579 } 1580 1581 void DevirtModule::exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1582 StringRef Name, uint32_t Const, 1583 uint32_t &Storage) { 1584 if (shouldExportConstantsAsAbsoluteSymbols()) { 1585 exportGlobal( 1586 Slot, Args, Name, 1587 ConstantExpr::getIntToPtr(ConstantInt::get(Int32Ty, Const), Int8PtrTy)); 1588 return; 1589 } 1590 1591 Storage = Const; 1592 } 1593 1594 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1595 StringRef Name) { 1596 Constant *C = 1597 M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Arr0Ty); 1598 auto *GV = dyn_cast<GlobalVariable>(C); 1599 if (GV) 1600 GV->setVisibility(GlobalValue::HiddenVisibility); 1601 return C; 1602 } 1603 1604 Constant *DevirtModule::importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1605 StringRef Name, IntegerType *IntTy, 1606 uint32_t Storage) { 1607 if (!shouldExportConstantsAsAbsoluteSymbols()) 1608 return ConstantInt::get(IntTy, Storage); 1609 1610 Constant *C = importGlobal(Slot, Args, Name); 1611 auto *GV = cast<GlobalVariable>(C->stripPointerCasts()); 1612 C = ConstantExpr::getPtrToInt(C, IntTy); 1613 1614 // We only need to set metadata if the global is newly created, in which 1615 // case it would not have hidden visibility. 1616 if (GV->hasMetadata(LLVMContext::MD_absolute_symbol)) 1617 return C; 1618 1619 auto SetAbsRange = [&](uint64_t Min, uint64_t Max) { 1620 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min)); 1621 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max)); 1622 GV->setMetadata(LLVMContext::MD_absolute_symbol, 1623 MDNode::get(M.getContext(), {MinC, MaxC})); 1624 }; 1625 unsigned AbsWidth = IntTy->getBitWidth(); 1626 if (AbsWidth == IntPtrTy->getBitWidth()) 1627 SetAbsRange(~0ull, ~0ull); // Full set. 1628 else 1629 SetAbsRange(0, 1ull << AbsWidth); 1630 return C; 1631 } 1632 1633 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1634 bool IsOne, 1635 Constant *UniqueMemberAddr) { 1636 for (auto &&Call : CSInfo.CallSites) { 1637 if (!OptimizedCalls.insert(&Call.CB).second) 1638 continue; 1639 IRBuilder<> B(&Call.CB); 1640 Value *Cmp = 1641 B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE, Call.VTable, 1642 B.CreateBitCast(UniqueMemberAddr, Call.VTable->getType())); 1643 Cmp = B.CreateZExt(Cmp, Call.CB.getType()); 1644 NumUniqueRetVal++; 1645 Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter, 1646 Cmp); 1647 } 1648 CSInfo.markDevirt(); 1649 } 1650 1651 Constant *DevirtModule::getMemberAddr(const TypeMemberInfo *M) { 1652 Constant *C = ConstantExpr::getBitCast(M->Bits->GV, Int8PtrTy); 1653 return ConstantExpr::getGetElementPtr(Int8Ty, C, 1654 ConstantInt::get(Int64Ty, M->Offset)); 1655 } 1656 1657 bool DevirtModule::tryUniqueRetValOpt( 1658 unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1659 CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res, 1660 VTableSlot Slot, ArrayRef<uint64_t> Args) { 1661 // IsOne controls whether we look for a 0 or a 1. 1662 auto tryUniqueRetValOptFor = [&](bool IsOne) { 1663 const TypeMemberInfo *UniqueMember = nullptr; 1664 for (const VirtualCallTarget &Target : TargetsForSlot) { 1665 if (Target.RetVal == (IsOne ? 1 : 0)) { 1666 if (UniqueMember) 1667 return false; 1668 UniqueMember = Target.TM; 1669 } 1670 } 1671 1672 // We should have found a unique member or bailed out by now. We already 1673 // checked for a uniform return value in tryUniformRetValOpt. 1674 assert(UniqueMember); 1675 1676 Constant *UniqueMemberAddr = getMemberAddr(UniqueMember); 1677 if (CSInfo.isExported()) { 1678 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniqueRetVal; 1679 Res->Info = IsOne; 1680 1681 exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr); 1682 } 1683 1684 // Replace each call with the comparison. 1685 applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne, 1686 UniqueMemberAddr); 1687 1688 // Update devirtualization statistics for targets. 1689 if (RemarksEnabled || AreStatisticsEnabled()) 1690 for (auto &&Target : TargetsForSlot) 1691 Target.WasDevirt = true; 1692 1693 return true; 1694 }; 1695 1696 if (BitWidth == 1) { 1697 if (tryUniqueRetValOptFor(true)) 1698 return true; 1699 if (tryUniqueRetValOptFor(false)) 1700 return true; 1701 } 1702 return false; 1703 } 1704 1705 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 1706 Constant *Byte, Constant *Bit) { 1707 for (auto Call : CSInfo.CallSites) { 1708 if (!OptimizedCalls.insert(&Call.CB).second) 1709 continue; 1710 auto *RetType = cast<IntegerType>(Call.CB.getType()); 1711 IRBuilder<> B(&Call.CB); 1712 Value *Addr = 1713 B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte); 1714 if (RetType->getBitWidth() == 1) { 1715 Value *Bits = B.CreateLoad(Int8Ty, Addr); 1716 Value *BitsAndBit = B.CreateAnd(Bits, Bit); 1717 auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0)); 1718 NumVirtConstProp1Bit++; 1719 Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled, 1720 OREGetter, IsBitSet); 1721 } else { 1722 Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo()); 1723 Value *Val = B.CreateLoad(RetType, ValAddr); 1724 NumVirtConstProp++; 1725 Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled, 1726 OREGetter, Val); 1727 } 1728 } 1729 CSInfo.markDevirt(); 1730 } 1731 1732 bool DevirtModule::tryVirtualConstProp( 1733 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1734 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1735 // This only works if the function returns an integer. 1736 auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType()); 1737 if (!RetType) 1738 return false; 1739 unsigned BitWidth = RetType->getBitWidth(); 1740 if (BitWidth > 64) 1741 return false; 1742 1743 // Make sure that each function is defined, does not access memory, takes at 1744 // least one argument, does not use its first argument (which we assume is 1745 // 'this'), and has the same return type. 1746 // 1747 // Note that we test whether this copy of the function is readnone, rather 1748 // than testing function attributes, which must hold for any copy of the 1749 // function, even a less optimized version substituted at link time. This is 1750 // sound because the virtual constant propagation optimizations effectively 1751 // inline all implementations of the virtual function into each call site, 1752 // rather than using function attributes to perform local optimization. 1753 for (VirtualCallTarget &Target : TargetsForSlot) { 1754 if (Target.Fn->isDeclaration() || 1755 computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) != 1756 FMRB_DoesNotAccessMemory || 1757 Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() || 1758 Target.Fn->getReturnType() != RetType) 1759 return false; 1760 } 1761 1762 for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) { 1763 if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first)) 1764 continue; 1765 1766 WholeProgramDevirtResolution::ByArg *ResByArg = nullptr; 1767 if (Res) 1768 ResByArg = &Res->ResByArg[CSByConstantArg.first]; 1769 1770 if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg)) 1771 continue; 1772 1773 if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second, 1774 ResByArg, Slot, CSByConstantArg.first)) 1775 continue; 1776 1777 // Find an allocation offset in bits in all vtables associated with the 1778 // type. 1779 uint64_t AllocBefore = 1780 findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth); 1781 uint64_t AllocAfter = 1782 findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth); 1783 1784 // Calculate the total amount of padding needed to store a value at both 1785 // ends of the object. 1786 uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0; 1787 for (auto &&Target : TargetsForSlot) { 1788 TotalPaddingBefore += std::max<int64_t>( 1789 (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0); 1790 TotalPaddingAfter += std::max<int64_t>( 1791 (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0); 1792 } 1793 1794 // If the amount of padding is too large, give up. 1795 // FIXME: do something smarter here. 1796 if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128) 1797 continue; 1798 1799 // Calculate the offset to the value as a (possibly negative) byte offset 1800 // and (if applicable) a bit offset, and store the values in the targets. 1801 int64_t OffsetByte; 1802 uint64_t OffsetBit; 1803 if (TotalPaddingBefore <= TotalPaddingAfter) 1804 setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte, 1805 OffsetBit); 1806 else 1807 setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte, 1808 OffsetBit); 1809 1810 if (RemarksEnabled || AreStatisticsEnabled()) 1811 for (auto &&Target : TargetsForSlot) 1812 Target.WasDevirt = true; 1813 1814 1815 if (CSByConstantArg.second.isExported()) { 1816 ResByArg->TheKind = WholeProgramDevirtResolution::ByArg::VirtualConstProp; 1817 exportConstant(Slot, CSByConstantArg.first, "byte", OffsetByte, 1818 ResByArg->Byte); 1819 exportConstant(Slot, CSByConstantArg.first, "bit", 1ULL << OffsetBit, 1820 ResByArg->Bit); 1821 } 1822 1823 // Rewrite each call to a load from OffsetByte/OffsetBit. 1824 Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte); 1825 Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit); 1826 applyVirtualConstProp(CSByConstantArg.second, 1827 TargetsForSlot[0].Fn->getName(), ByteConst, BitConst); 1828 } 1829 return true; 1830 } 1831 1832 void DevirtModule::rebuildGlobal(VTableBits &B) { 1833 if (B.Before.Bytes.empty() && B.After.Bytes.empty()) 1834 return; 1835 1836 // Align the before byte array to the global's minimum alignment so that we 1837 // don't break any alignment requirements on the global. 1838 Align Alignment = M.getDataLayout().getValueOrABITypeAlignment( 1839 B.GV->getAlign(), B.GV->getValueType()); 1840 B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), Alignment)); 1841 1842 // Before was stored in reverse order; flip it now. 1843 for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I) 1844 std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]); 1845 1846 // Build an anonymous global containing the before bytes, followed by the 1847 // original initializer, followed by the after bytes. 1848 auto NewInit = ConstantStruct::getAnon( 1849 {ConstantDataArray::get(M.getContext(), B.Before.Bytes), 1850 B.GV->getInitializer(), 1851 ConstantDataArray::get(M.getContext(), B.After.Bytes)}); 1852 auto NewGV = 1853 new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(), 1854 GlobalVariable::PrivateLinkage, NewInit, "", B.GV); 1855 NewGV->setSection(B.GV->getSection()); 1856 NewGV->setComdat(B.GV->getComdat()); 1857 NewGV->setAlignment(B.GV->getAlign()); 1858 1859 // Copy the original vtable's metadata to the anonymous global, adjusting 1860 // offsets as required. 1861 NewGV->copyMetadata(B.GV, B.Before.Bytes.size()); 1862 1863 // Build an alias named after the original global, pointing at the second 1864 // element (the original initializer). 1865 auto Alias = GlobalAlias::create( 1866 B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "", 1867 ConstantExpr::getGetElementPtr( 1868 NewInit->getType(), NewGV, 1869 ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0), 1870 ConstantInt::get(Int32Ty, 1)}), 1871 &M); 1872 Alias->setVisibility(B.GV->getVisibility()); 1873 Alias->takeName(B.GV); 1874 1875 B.GV->replaceAllUsesWith(Alias); 1876 B.GV->eraseFromParent(); 1877 } 1878 1879 bool DevirtModule::areRemarksEnabled() { 1880 const auto &FL = M.getFunctionList(); 1881 for (const Function &Fn : FL) { 1882 const auto &BBL = Fn.getBasicBlockList(); 1883 if (BBL.empty()) 1884 continue; 1885 auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front()); 1886 return DI.isEnabled(); 1887 } 1888 return false; 1889 } 1890 1891 void DevirtModule::scanTypeTestUsers( 1892 Function *TypeTestFunc, 1893 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 1894 // Find all virtual calls via a virtual table pointer %p under an assumption 1895 // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p 1896 // points to a member of the type identifier %md. Group calls by (type ID, 1897 // offset) pair (effectively the identity of the virtual function) and store 1898 // to CallSlots. 1899 for (Use &U : llvm::make_early_inc_range(TypeTestFunc->uses())) { 1900 auto *CI = dyn_cast<CallInst>(U.getUser()); 1901 if (!CI) 1902 continue; 1903 1904 // Search for virtual calls based on %p and add them to DevirtCalls. 1905 SmallVector<DevirtCallSite, 1> DevirtCalls; 1906 SmallVector<CallInst *, 1> Assumes; 1907 auto &DT = LookupDomTree(*CI->getFunction()); 1908 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT); 1909 1910 Metadata *TypeId = 1911 cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata(); 1912 // If we found any, add them to CallSlots. 1913 if (!Assumes.empty()) { 1914 Value *Ptr = CI->getArgOperand(0)->stripPointerCasts(); 1915 for (DevirtCallSite Call : DevirtCalls) 1916 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, nullptr); 1917 } 1918 1919 auto RemoveTypeTestAssumes = [&]() { 1920 // We no longer need the assumes or the type test. 1921 for (auto Assume : Assumes) 1922 Assume->eraseFromParent(); 1923 // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we 1924 // may use the vtable argument later. 1925 if (CI->use_empty()) 1926 CI->eraseFromParent(); 1927 }; 1928 1929 // At this point we could remove all type test assume sequences, as they 1930 // were originally inserted for WPD. However, we can keep these in the 1931 // code stream for later analysis (e.g. to help drive more efficient ICP 1932 // sequences). They will eventually be removed by a second LowerTypeTests 1933 // invocation that cleans them up. In order to do this correctly, the first 1934 // LowerTypeTests invocation needs to know that they have "Unknown" type 1935 // test resolution, so that they aren't treated as Unsat and lowered to 1936 // False, which will break any uses on assumes. Below we remove any type 1937 // test assumes that will not be treated as Unknown by LTT. 1938 1939 // The type test assumes will be treated by LTT as Unsat if the type id is 1940 // not used on a global (in which case it has no entry in the TypeIdMap). 1941 if (!TypeIdMap.count(TypeId)) 1942 RemoveTypeTestAssumes(); 1943 1944 // For ThinLTO importing, we need to remove the type test assumes if this is 1945 // an MDString type id without a corresponding TypeIdSummary. Any 1946 // non-MDString type ids are ignored and treated as Unknown by LTT, so their 1947 // type test assumes can be kept. If the MDString type id is missing a 1948 // TypeIdSummary (e.g. because there was no use on a vcall, preventing the 1949 // exporting phase of WPD from analyzing it), then it would be treated as 1950 // Unsat by LTT and we need to remove its type test assumes here. If not 1951 // used on a vcall we don't need them for later optimization use in any 1952 // case. 1953 else if (ImportSummary && isa<MDString>(TypeId)) { 1954 const TypeIdSummary *TidSummary = 1955 ImportSummary->getTypeIdSummary(cast<MDString>(TypeId)->getString()); 1956 if (!TidSummary) 1957 RemoveTypeTestAssumes(); 1958 else 1959 // If one was created it should not be Unsat, because if we reached here 1960 // the type id was used on a global. 1961 assert(TidSummary->TTRes.TheKind != TypeTestResolution::Unsat); 1962 } 1963 } 1964 } 1965 1966 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) { 1967 Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test); 1968 1969 for (Use &U : llvm::make_early_inc_range(TypeCheckedLoadFunc->uses())) { 1970 auto *CI = dyn_cast<CallInst>(U.getUser()); 1971 if (!CI) 1972 continue; 1973 1974 Value *Ptr = CI->getArgOperand(0); 1975 Value *Offset = CI->getArgOperand(1); 1976 Value *TypeIdValue = CI->getArgOperand(2); 1977 Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata(); 1978 1979 SmallVector<DevirtCallSite, 1> DevirtCalls; 1980 SmallVector<Instruction *, 1> LoadedPtrs; 1981 SmallVector<Instruction *, 1> Preds; 1982 bool HasNonCallUses = false; 1983 auto &DT = LookupDomTree(*CI->getFunction()); 1984 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds, 1985 HasNonCallUses, CI, DT); 1986 1987 // Start by generating "pessimistic" code that explicitly loads the function 1988 // pointer from the vtable and performs the type check. If possible, we will 1989 // eliminate the load and the type check later. 1990 1991 // If possible, only generate the load at the point where it is used. 1992 // This helps avoid unnecessary spills. 1993 IRBuilder<> LoadB( 1994 (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI); 1995 Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset); 1996 Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy)); 1997 Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr); 1998 1999 for (Instruction *LoadedPtr : LoadedPtrs) { 2000 LoadedPtr->replaceAllUsesWith(LoadedValue); 2001 LoadedPtr->eraseFromParent(); 2002 } 2003 2004 // Likewise for the type test. 2005 IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI); 2006 CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue}); 2007 2008 for (Instruction *Pred : Preds) { 2009 Pred->replaceAllUsesWith(TypeTestCall); 2010 Pred->eraseFromParent(); 2011 } 2012 2013 // We have already erased any extractvalue instructions that refer to the 2014 // intrinsic call, but the intrinsic may have other non-extractvalue uses 2015 // (although this is unlikely). In that case, explicitly build a pair and 2016 // RAUW it. 2017 if (!CI->use_empty()) { 2018 Value *Pair = PoisonValue::get(CI->getType()); 2019 IRBuilder<> B(CI); 2020 Pair = B.CreateInsertValue(Pair, LoadedValue, {0}); 2021 Pair = B.CreateInsertValue(Pair, TypeTestCall, {1}); 2022 CI->replaceAllUsesWith(Pair); 2023 } 2024 2025 // The number of unsafe uses is initially the number of uses. 2026 auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall]; 2027 NumUnsafeUses = DevirtCalls.size(); 2028 2029 // If the function pointer has a non-call user, we cannot eliminate the type 2030 // check, as one of those users may eventually call the pointer. Increment 2031 // the unsafe use count to make sure it cannot reach zero. 2032 if (HasNonCallUses) 2033 ++NumUnsafeUses; 2034 for (DevirtCallSite Call : DevirtCalls) { 2035 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CB, 2036 &NumUnsafeUses); 2037 } 2038 2039 CI->eraseFromParent(); 2040 } 2041 } 2042 2043 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) { 2044 auto *TypeId = dyn_cast<MDString>(Slot.TypeID); 2045 if (!TypeId) 2046 return; 2047 const TypeIdSummary *TidSummary = 2048 ImportSummary->getTypeIdSummary(TypeId->getString()); 2049 if (!TidSummary) 2050 return; 2051 auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset); 2052 if (ResI == TidSummary->WPDRes.end()) 2053 return; 2054 const WholeProgramDevirtResolution &Res = ResI->second; 2055 2056 if (Res.TheKind == WholeProgramDevirtResolution::SingleImpl) { 2057 assert(!Res.SingleImplName.empty()); 2058 // The type of the function in the declaration is irrelevant because every 2059 // call site will cast it to the correct type. 2060 Constant *SingleImpl = 2061 cast<Constant>(M.getOrInsertFunction(Res.SingleImplName, 2062 Type::getVoidTy(M.getContext())) 2063 .getCallee()); 2064 2065 // This is the import phase so we should not be exporting anything. 2066 bool IsExported = false; 2067 applySingleImplDevirt(SlotInfo, SingleImpl, IsExported); 2068 assert(!IsExported); 2069 } 2070 2071 for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) { 2072 auto I = Res.ResByArg.find(CSByConstantArg.first); 2073 if (I == Res.ResByArg.end()) 2074 continue; 2075 auto &ResByArg = I->second; 2076 // FIXME: We should figure out what to do about the "function name" argument 2077 // to the apply* functions, as the function names are unavailable during the 2078 // importing phase. For now we just pass the empty string. This does not 2079 // impact correctness because the function names are just used for remarks. 2080 switch (ResByArg.TheKind) { 2081 case WholeProgramDevirtResolution::ByArg::UniformRetVal: 2082 applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info); 2083 break; 2084 case WholeProgramDevirtResolution::ByArg::UniqueRetVal: { 2085 Constant *UniqueMemberAddr = 2086 importGlobal(Slot, CSByConstantArg.first, "unique_member"); 2087 applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info, 2088 UniqueMemberAddr); 2089 break; 2090 } 2091 case WholeProgramDevirtResolution::ByArg::VirtualConstProp: { 2092 Constant *Byte = importConstant(Slot, CSByConstantArg.first, "byte", 2093 Int32Ty, ResByArg.Byte); 2094 Constant *Bit = importConstant(Slot, CSByConstantArg.first, "bit", Int8Ty, 2095 ResByArg.Bit); 2096 applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit); 2097 break; 2098 } 2099 default: 2100 break; 2101 } 2102 } 2103 2104 if (Res.TheKind == WholeProgramDevirtResolution::BranchFunnel) { 2105 // The type of the function is irrelevant, because it's bitcast at calls 2106 // anyhow. 2107 Constant *JT = cast<Constant>( 2108 M.getOrInsertFunction(getGlobalName(Slot, {}, "branch_funnel"), 2109 Type::getVoidTy(M.getContext())) 2110 .getCallee()); 2111 bool IsExported = false; 2112 applyICallBranchFunnel(SlotInfo, JT, IsExported); 2113 assert(!IsExported); 2114 } 2115 } 2116 2117 void DevirtModule::removeRedundantTypeTests() { 2118 auto True = ConstantInt::getTrue(M.getContext()); 2119 for (auto &&U : NumUnsafeUsesForTypeTest) { 2120 if (U.second == 0) { 2121 U.first->replaceAllUsesWith(True); 2122 U.first->eraseFromParent(); 2123 } 2124 } 2125 } 2126 2127 ValueInfo 2128 DevirtModule::lookUpFunctionValueInfo(Function *TheFn, 2129 ModuleSummaryIndex *ExportSummary) { 2130 assert((ExportSummary != nullptr) && 2131 "Caller guarantees ExportSummary is not nullptr"); 2132 2133 const auto TheFnGUID = TheFn->getGUID(); 2134 const auto TheFnGUIDWithExportedName = GlobalValue::getGUID(TheFn->getName()); 2135 // Look up ValueInfo with the GUID in the current linkage. 2136 ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFnGUID); 2137 // If no entry is found and GUID is different from GUID computed using 2138 // exported name, look up ValueInfo with the exported name unconditionally. 2139 // This is a fallback. 2140 // 2141 // The reason to have a fallback: 2142 // 1. LTO could enable global value internalization via 2143 // `enable-lto-internalization`. 2144 // 2. The GUID in ExportedSummary is computed using exported name. 2145 if ((!TheFnVI) && (TheFnGUID != TheFnGUIDWithExportedName)) { 2146 TheFnVI = ExportSummary->getValueInfo(TheFnGUIDWithExportedName); 2147 } 2148 return TheFnVI; 2149 } 2150 2151 bool DevirtModule::mustBeUnreachableFunction( 2152 Function *const F, ModuleSummaryIndex *ExportSummary) { 2153 // First, learn unreachability by analyzing function IR. 2154 if (!F->isDeclaration()) { 2155 // A function must be unreachable if its entry block ends with an 2156 // 'unreachable'. 2157 return isa<UnreachableInst>(F->getEntryBlock().getTerminator()); 2158 } 2159 // Learn unreachability from ExportSummary if ExportSummary is present. 2160 return ExportSummary && 2161 ::mustBeUnreachableFunction( 2162 DevirtModule::lookUpFunctionValueInfo(F, ExportSummary)); 2163 } 2164 2165 bool DevirtModule::run() { 2166 // If only some of the modules were split, we cannot correctly perform 2167 // this transformation. We already checked for the presense of type tests 2168 // with partially split modules during the thin link, and would have emitted 2169 // an error if any were found, so here we can simply return. 2170 if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) || 2171 (ImportSummary && ImportSummary->partiallySplitLTOUnits())) 2172 return false; 2173 2174 Function *TypeTestFunc = 2175 M.getFunction(Intrinsic::getName(Intrinsic::type_test)); 2176 Function *TypeCheckedLoadFunc = 2177 M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load)); 2178 Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume)); 2179 2180 // Normally if there are no users of the devirtualization intrinsics in the 2181 // module, this pass has nothing to do. But if we are exporting, we also need 2182 // to handle any users that appear only in the function summaries. 2183 if (!ExportSummary && 2184 (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc || 2185 AssumeFunc->use_empty()) && 2186 (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty())) 2187 return false; 2188 2189 // Rebuild type metadata into a map for easy lookup. 2190 std::vector<VTableBits> Bits; 2191 DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap; 2192 buildTypeIdentifierMap(Bits, TypeIdMap); 2193 2194 if (TypeTestFunc && AssumeFunc) 2195 scanTypeTestUsers(TypeTestFunc, TypeIdMap); 2196 2197 if (TypeCheckedLoadFunc) 2198 scanTypeCheckedLoadUsers(TypeCheckedLoadFunc); 2199 2200 if (ImportSummary) { 2201 for (auto &S : CallSlots) 2202 importResolution(S.first, S.second); 2203 2204 removeRedundantTypeTests(); 2205 2206 // We have lowered or deleted the type intrinsics, so we will no longer have 2207 // enough information to reason about the liveness of virtual function 2208 // pointers in GlobalDCE. 2209 for (GlobalVariable &GV : M.globals()) 2210 GV.eraseMetadata(LLVMContext::MD_vcall_visibility); 2211 2212 // The rest of the code is only necessary when exporting or during regular 2213 // LTO, so we are done. 2214 return true; 2215 } 2216 2217 if (TypeIdMap.empty()) 2218 return true; 2219 2220 // Collect information from summary about which calls to try to devirtualize. 2221 if (ExportSummary) { 2222 DenseMap<GlobalValue::GUID, TinyPtrVector<Metadata *>> MetadataByGUID; 2223 for (auto &P : TypeIdMap) { 2224 if (auto *TypeId = dyn_cast<MDString>(P.first)) 2225 MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back( 2226 TypeId); 2227 } 2228 2229 for (auto &P : *ExportSummary) { 2230 for (auto &S : P.second.SummaryList) { 2231 auto *FS = dyn_cast<FunctionSummary>(S.get()); 2232 if (!FS) 2233 continue; 2234 // FIXME: Only add live functions. 2235 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 2236 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 2237 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 2238 } 2239 } 2240 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 2241 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 2242 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 2243 } 2244 } 2245 for (const FunctionSummary::ConstVCall &VC : 2246 FS->type_test_assume_const_vcalls()) { 2247 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 2248 CallSlots[{MD, VC.VFunc.Offset}] 2249 .ConstCSInfo[VC.Args] 2250 .addSummaryTypeTestAssumeUser(FS); 2251 } 2252 } 2253 for (const FunctionSummary::ConstVCall &VC : 2254 FS->type_checked_load_const_vcalls()) { 2255 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 2256 CallSlots[{MD, VC.VFunc.Offset}] 2257 .ConstCSInfo[VC.Args] 2258 .addSummaryTypeCheckedLoadUser(FS); 2259 } 2260 } 2261 } 2262 } 2263 } 2264 2265 // For each (type, offset) pair: 2266 bool DidVirtualConstProp = false; 2267 std::map<std::string, Function*> DevirtTargets; 2268 for (auto &S : CallSlots) { 2269 // Search each of the members of the type identifier for the virtual 2270 // function implementation at offset S.first.ByteOffset, and add to 2271 // TargetsForSlot. 2272 std::vector<VirtualCallTarget> TargetsForSlot; 2273 WholeProgramDevirtResolution *Res = nullptr; 2274 const std::set<TypeMemberInfo> &TypeMemberInfos = TypeIdMap[S.first.TypeID]; 2275 if (ExportSummary && isa<MDString>(S.first.TypeID) && 2276 TypeMemberInfos.size()) 2277 // For any type id used on a global's type metadata, create the type id 2278 // summary resolution regardless of whether we can devirtualize, so that 2279 // lower type tests knows the type id is not Unsat. If it was not used on 2280 // a global's type metadata, the TypeIdMap entry set will be empty, and 2281 // we don't want to create an entry (with the default Unknown type 2282 // resolution), which can prevent detection of the Unsat. 2283 Res = &ExportSummary 2284 ->getOrInsertTypeIdSummary( 2285 cast<MDString>(S.first.TypeID)->getString()) 2286 .WPDRes[S.first.ByteOffset]; 2287 if (tryFindVirtualCallTargets(TargetsForSlot, TypeMemberInfos, 2288 S.first.ByteOffset, ExportSummary)) { 2289 2290 if (!trySingleImplDevirt(ExportSummary, TargetsForSlot, S.second, Res)) { 2291 DidVirtualConstProp |= 2292 tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first); 2293 2294 tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first); 2295 } 2296 2297 // Collect functions devirtualized at least for one call site for stats. 2298 if (RemarksEnabled || AreStatisticsEnabled()) 2299 for (const auto &T : TargetsForSlot) 2300 if (T.WasDevirt) 2301 DevirtTargets[std::string(T.Fn->getName())] = T.Fn; 2302 } 2303 2304 // CFI-specific: if we are exporting and any llvm.type.checked.load 2305 // intrinsics were *not* devirtualized, we need to add the resulting 2306 // llvm.type.test intrinsics to the function summaries so that the 2307 // LowerTypeTests pass will export them. 2308 if (ExportSummary && isa<MDString>(S.first.TypeID)) { 2309 auto GUID = 2310 GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString()); 2311 for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers) 2312 FS->addTypeTest(GUID); 2313 for (auto &CCS : S.second.ConstCSInfo) 2314 for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers) 2315 FS->addTypeTest(GUID); 2316 } 2317 } 2318 2319 if (RemarksEnabled) { 2320 // Generate remarks for each devirtualized function. 2321 for (const auto &DT : DevirtTargets) { 2322 Function *F = DT.second; 2323 2324 using namespace ore; 2325 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F) 2326 << "devirtualized " 2327 << NV("FunctionName", DT.first)); 2328 } 2329 } 2330 2331 NumDevirtTargets += DevirtTargets.size(); 2332 2333 removeRedundantTypeTests(); 2334 2335 // Rebuild each global we touched as part of virtual constant propagation to 2336 // include the before and after bytes. 2337 if (DidVirtualConstProp) 2338 for (VTableBits &B : Bits) 2339 rebuildGlobal(B); 2340 2341 // We have lowered or deleted the type intrinsics, so we will no longer have 2342 // enough information to reason about the liveness of virtual function 2343 // pointers in GlobalDCE. 2344 for (GlobalVariable &GV : M.globals()) 2345 GV.eraseMetadata(LLVMContext::MD_vcall_visibility); 2346 2347 return true; 2348 } 2349 2350 void DevirtIndex::run() { 2351 if (ExportSummary.typeIdCompatibleVtableMap().empty()) 2352 return; 2353 2354 DenseMap<GlobalValue::GUID, std::vector<StringRef>> NameByGUID; 2355 for (auto &P : ExportSummary.typeIdCompatibleVtableMap()) { 2356 NameByGUID[GlobalValue::getGUID(P.first)].push_back(P.first); 2357 } 2358 2359 // Collect information from summary about which calls to try to devirtualize. 2360 for (auto &P : ExportSummary) { 2361 for (auto &S : P.second.SummaryList) { 2362 auto *FS = dyn_cast<FunctionSummary>(S.get()); 2363 if (!FS) 2364 continue; 2365 // FIXME: Only add live functions. 2366 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 2367 for (StringRef Name : NameByGUID[VF.GUID]) { 2368 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 2369 } 2370 } 2371 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 2372 for (StringRef Name : NameByGUID[VF.GUID]) { 2373 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 2374 } 2375 } 2376 for (const FunctionSummary::ConstVCall &VC : 2377 FS->type_test_assume_const_vcalls()) { 2378 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 2379 CallSlots[{Name, VC.VFunc.Offset}] 2380 .ConstCSInfo[VC.Args] 2381 .addSummaryTypeTestAssumeUser(FS); 2382 } 2383 } 2384 for (const FunctionSummary::ConstVCall &VC : 2385 FS->type_checked_load_const_vcalls()) { 2386 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 2387 CallSlots[{Name, VC.VFunc.Offset}] 2388 .ConstCSInfo[VC.Args] 2389 .addSummaryTypeCheckedLoadUser(FS); 2390 } 2391 } 2392 } 2393 } 2394 2395 std::set<ValueInfo> DevirtTargets; 2396 // For each (type, offset) pair: 2397 for (auto &S : CallSlots) { 2398 // Search each of the members of the type identifier for the virtual 2399 // function implementation at offset S.first.ByteOffset, and add to 2400 // TargetsForSlot. 2401 std::vector<ValueInfo> TargetsForSlot; 2402 auto TidSummary = ExportSummary.getTypeIdCompatibleVtableSummary(S.first.TypeID); 2403 assert(TidSummary); 2404 // Create the type id summary resolution regardlness of whether we can 2405 // devirtualize, so that lower type tests knows the type id is used on 2406 // a global and not Unsat. 2407 WholeProgramDevirtResolution *Res = 2408 &ExportSummary.getOrInsertTypeIdSummary(S.first.TypeID) 2409 .WPDRes[S.first.ByteOffset]; 2410 if (tryFindVirtualCallTargets(TargetsForSlot, *TidSummary, 2411 S.first.ByteOffset)) { 2412 2413 if (!trySingleImplDevirt(TargetsForSlot, S.first, S.second, Res, 2414 DevirtTargets)) 2415 continue; 2416 } 2417 } 2418 2419 // Optionally have the thin link print message for each devirtualized 2420 // function. 2421 if (PrintSummaryDevirt) 2422 for (const auto &DT : DevirtTargets) 2423 errs() << "Devirtualized call to " << DT << "\n"; 2424 2425 NumDevirtTargets += DevirtTargets.size(); 2426 } 2427