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