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/Pass.h" 84 #include "llvm/PassRegistry.h" 85 #include "llvm/PassSupport.h" 86 #include "llvm/Support/Casting.h" 87 #include "llvm/Support/Error.h" 88 #include "llvm/Support/FileSystem.h" 89 #include "llvm/Support/MathExtras.h" 90 #include "llvm/Transforms/IPO.h" 91 #include "llvm/Transforms/IPO/FunctionAttrs.h" 92 #include "llvm/Transforms/Utils/Evaluator.h" 93 #include <algorithm> 94 #include <cstddef> 95 #include <map> 96 #include <set> 97 #include <string> 98 99 using namespace llvm; 100 using namespace wholeprogramdevirt; 101 102 #define DEBUG_TYPE "wholeprogramdevirt" 103 104 static cl::opt<PassSummaryAction> ClSummaryAction( 105 "wholeprogramdevirt-summary-action", 106 cl::desc("What to do with the summary when running this pass"), 107 cl::values(clEnumValN(PassSummaryAction::None, "none", "Do nothing"), 108 clEnumValN(PassSummaryAction::Import, "import", 109 "Import typeid resolutions from summary and globals"), 110 clEnumValN(PassSummaryAction::Export, "export", 111 "Export typeid resolutions to summary and globals")), 112 cl::Hidden); 113 114 static cl::opt<std::string> ClReadSummary( 115 "wholeprogramdevirt-read-summary", 116 cl::desc("Read summary from given YAML file before running pass"), 117 cl::Hidden); 118 119 static cl::opt<std::string> ClWriteSummary( 120 "wholeprogramdevirt-write-summary", 121 cl::desc("Write summary to given YAML file after running pass"), 122 cl::Hidden); 123 124 static cl::opt<unsigned> 125 ClThreshold("wholeprogramdevirt-branch-funnel-threshold", cl::Hidden, 126 cl::init(10), cl::ZeroOrMore, 127 cl::desc("Maximum number of call targets per " 128 "call site to enable branch funnels")); 129 130 static cl::opt<bool> 131 PrintSummaryDevirt("wholeprogramdevirt-print-index-based", cl::Hidden, 132 cl::init(false), cl::ZeroOrMore, 133 cl::desc("Print index-based devirtualization messages")); 134 135 // Find the minimum offset that we may store a value of size Size bits at. If 136 // IsAfter is set, look for an offset before the object, otherwise look for an 137 // offset after the object. 138 uint64_t 139 wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets, 140 bool IsAfter, uint64_t Size) { 141 // Find a minimum offset taking into account only vtable sizes. 142 uint64_t MinByte = 0; 143 for (const VirtualCallTarget &Target : Targets) { 144 if (IsAfter) 145 MinByte = std::max(MinByte, Target.minAfterBytes()); 146 else 147 MinByte = std::max(MinByte, Target.minBeforeBytes()); 148 } 149 150 // Build a vector of arrays of bytes covering, for each target, a slice of the 151 // used region (see AccumBitVector::BytesUsed in 152 // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively, 153 // this aligns the used regions to start at MinByte. 154 // 155 // In this example, A, B and C are vtables, # is a byte already allocated for 156 // a virtual function pointer, AAAA... (etc.) are the used regions for the 157 // vtables and Offset(X) is the value computed for the Offset variable below 158 // for X. 159 // 160 // Offset(A) 161 // | | 162 // |MinByte 163 // A: ################AAAAAAAA|AAAAAAAA 164 // B: ########BBBBBBBBBBBBBBBB|BBBB 165 // C: ########################|CCCCCCCCCCCCCCCC 166 // | Offset(B) | 167 // 168 // This code produces the slices of A, B and C that appear after the divider 169 // at MinByte. 170 std::vector<ArrayRef<uint8_t>> Used; 171 for (const VirtualCallTarget &Target : Targets) { 172 ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed 173 : Target.TM->Bits->Before.BytesUsed; 174 uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes() 175 : MinByte - Target.minBeforeBytes(); 176 177 // Disregard used regions that are smaller than Offset. These are 178 // effectively all-free regions that do not need to be checked. 179 if (VTUsed.size() > Offset) 180 Used.push_back(VTUsed.slice(Offset)); 181 } 182 183 if (Size == 1) { 184 // Find a free bit in each member of Used. 185 for (unsigned I = 0;; ++I) { 186 uint8_t BitsUsed = 0; 187 for (auto &&B : Used) 188 if (I < B.size()) 189 BitsUsed |= B[I]; 190 if (BitsUsed != 0xff) 191 return (MinByte + I) * 8 + 192 countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined); 193 } 194 } else { 195 // Find a free (Size/8) byte region in each member of Used. 196 // FIXME: see if alignment helps. 197 for (unsigned I = 0;; ++I) { 198 for (auto &&B : Used) { 199 unsigned Byte = 0; 200 while ((I + Byte) < B.size() && Byte < (Size / 8)) { 201 if (B[I + Byte]) 202 goto NextI; 203 ++Byte; 204 } 205 } 206 return (MinByte + I) * 8; 207 NextI:; 208 } 209 } 210 } 211 212 void wholeprogramdevirt::setBeforeReturnValues( 213 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore, 214 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 215 if (BitWidth == 1) 216 OffsetByte = -(AllocBefore / 8 + 1); 217 else 218 OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8); 219 OffsetBit = AllocBefore % 8; 220 221 for (VirtualCallTarget &Target : Targets) { 222 if (BitWidth == 1) 223 Target.setBeforeBit(AllocBefore); 224 else 225 Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8); 226 } 227 } 228 229 void wholeprogramdevirt::setAfterReturnValues( 230 MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter, 231 unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) { 232 if (BitWidth == 1) 233 OffsetByte = AllocAfter / 8; 234 else 235 OffsetByte = (AllocAfter + 7) / 8; 236 OffsetBit = AllocAfter % 8; 237 238 for (VirtualCallTarget &Target : Targets) { 239 if (BitWidth == 1) 240 Target.setAfterBit(AllocAfter); 241 else 242 Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8); 243 } 244 } 245 246 VirtualCallTarget::VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM) 247 : Fn(Fn), TM(TM), 248 IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {} 249 250 namespace { 251 252 // A slot in a set of virtual tables. The TypeID identifies the set of virtual 253 // tables, and the ByteOffset is the offset in bytes from the address point to 254 // the virtual function pointer. 255 struct VTableSlot { 256 Metadata *TypeID; 257 uint64_t ByteOffset; 258 }; 259 260 } // end anonymous namespace 261 262 namespace llvm { 263 264 template <> struct DenseMapInfo<VTableSlot> { 265 static VTableSlot getEmptyKey() { 266 return {DenseMapInfo<Metadata *>::getEmptyKey(), 267 DenseMapInfo<uint64_t>::getEmptyKey()}; 268 } 269 static VTableSlot getTombstoneKey() { 270 return {DenseMapInfo<Metadata *>::getTombstoneKey(), 271 DenseMapInfo<uint64_t>::getTombstoneKey()}; 272 } 273 static unsigned getHashValue(const VTableSlot &I) { 274 return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^ 275 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 276 } 277 static bool isEqual(const VTableSlot &LHS, 278 const VTableSlot &RHS) { 279 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 280 } 281 }; 282 283 template <> struct DenseMapInfo<VTableSlotSummary> { 284 static VTableSlotSummary getEmptyKey() { 285 return {DenseMapInfo<StringRef>::getEmptyKey(), 286 DenseMapInfo<uint64_t>::getEmptyKey()}; 287 } 288 static VTableSlotSummary getTombstoneKey() { 289 return {DenseMapInfo<StringRef>::getTombstoneKey(), 290 DenseMapInfo<uint64_t>::getTombstoneKey()}; 291 } 292 static unsigned getHashValue(const VTableSlotSummary &I) { 293 return DenseMapInfo<StringRef>::getHashValue(I.TypeID) ^ 294 DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset); 295 } 296 static bool isEqual(const VTableSlotSummary &LHS, 297 const VTableSlotSummary &RHS) { 298 return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset; 299 } 300 }; 301 302 } // end namespace llvm 303 304 namespace { 305 306 // A virtual call site. VTable is the loaded virtual table pointer, and CS is 307 // the indirect virtual call. 308 struct VirtualCallSite { 309 Value *VTable; 310 CallSite CS; 311 312 // If non-null, this field points to the associated unsafe use count stored in 313 // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description 314 // of that field for details. 315 unsigned *NumUnsafeUses; 316 317 void 318 emitRemark(const StringRef OptName, const StringRef TargetName, 319 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) { 320 Function *F = CS.getCaller(); 321 DebugLoc DLoc = CS->getDebugLoc(); 322 BasicBlock *Block = CS.getParent(); 323 324 using namespace ore; 325 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, OptName, DLoc, Block) 326 << NV("Optimization", OptName) 327 << ": devirtualized a call to " 328 << NV("FunctionName", TargetName)); 329 } 330 331 void replaceAndErase( 332 const StringRef OptName, const StringRef TargetName, bool RemarksEnabled, 333 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 334 Value *New) { 335 if (RemarksEnabled) 336 emitRemark(OptName, TargetName, OREGetter); 337 CS->replaceAllUsesWith(New); 338 if (auto II = dyn_cast<InvokeInst>(CS.getInstruction())) { 339 BranchInst::Create(II->getNormalDest(), CS.getInstruction()); 340 II->getUnwindDest()->removePredecessor(II->getParent()); 341 } 342 CS->eraseFromParent(); 343 // This use is no longer unsafe. 344 if (NumUnsafeUses) 345 --*NumUnsafeUses; 346 } 347 }; 348 349 // Call site information collected for a specific VTableSlot and possibly a list 350 // of constant integer arguments. The grouping by arguments is handled by the 351 // VTableSlotInfo class. 352 struct CallSiteInfo { 353 /// The set of call sites for this slot. Used during regular LTO and the 354 /// import phase of ThinLTO (as well as the export phase of ThinLTO for any 355 /// call sites that appear in the merged module itself); in each of these 356 /// cases we are directly operating on the call sites at the IR level. 357 std::vector<VirtualCallSite> CallSites; 358 359 /// Whether all call sites represented by this CallSiteInfo, including those 360 /// in summaries, have been devirtualized. This starts off as true because a 361 /// default constructed CallSiteInfo represents no call sites. 362 bool AllCallSitesDevirted = true; 363 364 // These fields are used during the export phase of ThinLTO and reflect 365 // information collected from function summaries. 366 367 /// Whether any function summary contains an llvm.assume(llvm.type.test) for 368 /// this slot. 369 bool SummaryHasTypeTestAssumeUsers = false; 370 371 /// CFI-specific: a vector containing the list of function summaries that use 372 /// the llvm.type.checked.load intrinsic and therefore will require 373 /// resolutions for llvm.type.test in order to implement CFI checks if 374 /// devirtualization was unsuccessful. If devirtualization was successful, the 375 /// pass will clear this vector by calling markDevirt(). If at the end of the 376 /// pass the vector is non-empty, we will need to add a use of llvm.type.test 377 /// to each of the function summaries in the vector. 378 std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers; 379 std::vector<FunctionSummary *> SummaryTypeTestAssumeUsers; 380 381 bool isExported() const { 382 return SummaryHasTypeTestAssumeUsers || 383 !SummaryTypeCheckedLoadUsers.empty(); 384 } 385 386 void markSummaryHasTypeTestAssumeUsers() { 387 SummaryHasTypeTestAssumeUsers = true; 388 AllCallSitesDevirted = false; 389 } 390 391 void addSummaryTypeCheckedLoadUser(FunctionSummary *FS) { 392 SummaryTypeCheckedLoadUsers.push_back(FS); 393 AllCallSitesDevirted = false; 394 } 395 396 void addSummaryTypeTestAssumeUser(FunctionSummary *FS) { 397 SummaryTypeTestAssumeUsers.push_back(FS); 398 markSummaryHasTypeTestAssumeUsers(); 399 } 400 401 void markDevirt() { 402 AllCallSitesDevirted = true; 403 404 // As explained in the comment for SummaryTypeCheckedLoadUsers. 405 SummaryTypeCheckedLoadUsers.clear(); 406 } 407 }; 408 409 // Call site information collected for a specific VTableSlot. 410 struct VTableSlotInfo { 411 // The set of call sites which do not have all constant integer arguments 412 // (excluding "this"). 413 CallSiteInfo CSInfo; 414 415 // The set of call sites with all constant integer arguments (excluding 416 // "this"), grouped by argument list. 417 std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo; 418 419 void addCallSite(Value *VTable, CallSite CS, unsigned *NumUnsafeUses); 420 421 private: 422 CallSiteInfo &findCallSiteInfo(CallSite CS); 423 }; 424 425 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallSite CS) { 426 std::vector<uint64_t> Args; 427 auto *CI = dyn_cast<IntegerType>(CS.getType()); 428 if (!CI || CI->getBitWidth() > 64 || CS.arg_empty()) 429 return CSInfo; 430 for (auto &&Arg : make_range(CS.arg_begin() + 1, CS.arg_end())) { 431 auto *CI = dyn_cast<ConstantInt>(Arg); 432 if (!CI || CI->getBitWidth() > 64) 433 return CSInfo; 434 Args.push_back(CI->getZExtValue()); 435 } 436 return ConstCSInfo[Args]; 437 } 438 439 void VTableSlotInfo::addCallSite(Value *VTable, CallSite CS, 440 unsigned *NumUnsafeUses) { 441 auto &CSI = findCallSiteInfo(CS); 442 CSI.AllCallSitesDevirted = false; 443 CSI.CallSites.push_back({VTable, CS, NumUnsafeUses}); 444 } 445 446 struct DevirtModule { 447 Module &M; 448 function_ref<AAResults &(Function &)> AARGetter; 449 function_ref<DominatorTree &(Function &)> LookupDomTree; 450 451 ModuleSummaryIndex *ExportSummary; 452 const ModuleSummaryIndex *ImportSummary; 453 454 IntegerType *Int8Ty; 455 PointerType *Int8PtrTy; 456 IntegerType *Int32Ty; 457 IntegerType *Int64Ty; 458 IntegerType *IntPtrTy; 459 460 bool RemarksEnabled; 461 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter; 462 463 MapVector<VTableSlot, VTableSlotInfo> CallSlots; 464 465 // This map keeps track of the number of "unsafe" uses of a loaded function 466 // pointer. The key is the associated llvm.type.test intrinsic call generated 467 // by this pass. An unsafe use is one that calls the loaded function pointer 468 // directly. Every time we eliminate an unsafe use (for example, by 469 // devirtualizing it or by applying virtual constant propagation), we 470 // decrement the value stored in this map. If a value reaches zero, we can 471 // eliminate the type check by RAUWing the associated llvm.type.test call with 472 // true. 473 std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest; 474 475 DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter, 476 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 477 function_ref<DominatorTree &(Function &)> LookupDomTree, 478 ModuleSummaryIndex *ExportSummary, 479 const ModuleSummaryIndex *ImportSummary) 480 : M(M), AARGetter(AARGetter), LookupDomTree(LookupDomTree), 481 ExportSummary(ExportSummary), ImportSummary(ImportSummary), 482 Int8Ty(Type::getInt8Ty(M.getContext())), 483 Int8PtrTy(Type::getInt8PtrTy(M.getContext())), 484 Int32Ty(Type::getInt32Ty(M.getContext())), 485 Int64Ty(Type::getInt64Ty(M.getContext())), 486 IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)), 487 RemarksEnabled(areRemarksEnabled()), OREGetter(OREGetter) { 488 assert(!(ExportSummary && ImportSummary)); 489 } 490 491 bool areRemarksEnabled(); 492 493 void scanTypeTestUsers(Function *TypeTestFunc, Function *AssumeFunc); 494 void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc); 495 496 void buildTypeIdentifierMap( 497 std::vector<VTableBits> &Bits, 498 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap); 499 Constant *getPointerAtOffset(Constant *I, uint64_t Offset); 500 bool 501 tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot, 502 const std::set<TypeMemberInfo> &TypeMemberInfos, 503 uint64_t ByteOffset); 504 505 void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn, 506 bool &IsExported); 507 bool trySingleImplDevirt(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 508 VTableSlotInfo &SlotInfo, 509 WholeProgramDevirtResolution *Res); 510 511 void applyICallBranchFunnel(VTableSlotInfo &SlotInfo, Constant *JT, 512 bool &IsExported); 513 void tryICallBranchFunnel(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 514 VTableSlotInfo &SlotInfo, 515 WholeProgramDevirtResolution *Res, VTableSlot Slot); 516 517 bool tryEvaluateFunctionsWithArgs( 518 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 519 ArrayRef<uint64_t> Args); 520 521 void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 522 uint64_t TheRetVal); 523 bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 524 CallSiteInfo &CSInfo, 525 WholeProgramDevirtResolution::ByArg *Res); 526 527 // Returns the global symbol name that is used to export information about the 528 // given vtable slot and list of arguments. 529 std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args, 530 StringRef Name); 531 532 bool shouldExportConstantsAsAbsoluteSymbols(); 533 534 // This function is called during the export phase to create a symbol 535 // definition containing information about the given vtable slot and list of 536 // arguments. 537 void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 538 Constant *C); 539 void exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name, 540 uint32_t Const, uint32_t &Storage); 541 542 // This function is called during the import phase to create a reference to 543 // the symbol definition created during the export phase. 544 Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 545 StringRef Name); 546 Constant *importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 547 StringRef Name, IntegerType *IntTy, 548 uint32_t Storage); 549 550 Constant *getMemberAddr(const TypeMemberInfo *M); 551 552 void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne, 553 Constant *UniqueMemberAddr); 554 bool tryUniqueRetValOpt(unsigned BitWidth, 555 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 556 CallSiteInfo &CSInfo, 557 WholeProgramDevirtResolution::ByArg *Res, 558 VTableSlot Slot, ArrayRef<uint64_t> Args); 559 560 void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 561 Constant *Byte, Constant *Bit); 562 bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot, 563 VTableSlotInfo &SlotInfo, 564 WholeProgramDevirtResolution *Res, VTableSlot Slot); 565 566 void rebuildGlobal(VTableBits &B); 567 568 // Apply the summary resolution for Slot to all virtual calls in SlotInfo. 569 void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo); 570 571 // If we were able to eliminate all unsafe uses for a type checked load, 572 // eliminate the associated type tests by replacing them with true. 573 void removeRedundantTypeTests(); 574 575 bool run(); 576 577 // Lower the module using the action and summary passed as command line 578 // arguments. For testing purposes only. 579 static bool 580 runForTesting(Module &M, function_ref<AAResults &(Function &)> AARGetter, 581 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 582 function_ref<DominatorTree &(Function &)> LookupDomTree); 583 }; 584 585 struct DevirtIndex { 586 ModuleSummaryIndex &ExportSummary; 587 // The set in which to record GUIDs exported from their module by 588 // devirtualization, used by client to ensure they are not internalized. 589 std::set<GlobalValue::GUID> &ExportedGUIDs; 590 // A map in which to record the information necessary to locate the WPD 591 // resolution for local targets in case they are exported by cross module 592 // importing. 593 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap; 594 595 MapVector<VTableSlotSummary, VTableSlotInfo> CallSlots; 596 597 DevirtIndex( 598 ModuleSummaryIndex &ExportSummary, 599 std::set<GlobalValue::GUID> &ExportedGUIDs, 600 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) 601 : ExportSummary(ExportSummary), ExportedGUIDs(ExportedGUIDs), 602 LocalWPDTargetsMap(LocalWPDTargetsMap) {} 603 604 bool tryFindVirtualCallTargets(std::vector<ValueInfo> &TargetsForSlot, 605 const TypeIdCompatibleVtableInfo TIdInfo, 606 uint64_t ByteOffset); 607 608 bool trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 609 VTableSlotSummary &SlotSummary, 610 VTableSlotInfo &SlotInfo, 611 WholeProgramDevirtResolution *Res, 612 std::set<ValueInfo> &DevirtTargets); 613 614 void run(); 615 }; 616 617 struct WholeProgramDevirt : public ModulePass { 618 static char ID; 619 620 bool UseCommandLine = false; 621 622 ModuleSummaryIndex *ExportSummary; 623 const ModuleSummaryIndex *ImportSummary; 624 625 WholeProgramDevirt() : ModulePass(ID), UseCommandLine(true) { 626 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry()); 627 } 628 629 WholeProgramDevirt(ModuleSummaryIndex *ExportSummary, 630 const ModuleSummaryIndex *ImportSummary) 631 : ModulePass(ID), ExportSummary(ExportSummary), 632 ImportSummary(ImportSummary) { 633 initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry()); 634 } 635 636 bool runOnModule(Module &M) override { 637 if (skipModule(M)) 638 return false; 639 640 // In the new pass manager, we can request the optimization 641 // remark emitter pass on a per-function-basis, which the 642 // OREGetter will do for us. 643 // In the old pass manager, this is harder, so we just build 644 // an optimization remark emitter on the fly, when we need it. 645 std::unique_ptr<OptimizationRemarkEmitter> ORE; 646 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { 647 ORE = make_unique<OptimizationRemarkEmitter>(F); 648 return *ORE; 649 }; 650 651 auto LookupDomTree = [this](Function &F) -> DominatorTree & { 652 return this->getAnalysis<DominatorTreeWrapperPass>(F).getDomTree(); 653 }; 654 655 if (UseCommandLine) 656 return DevirtModule::runForTesting(M, LegacyAARGetter(*this), OREGetter, 657 LookupDomTree); 658 659 return DevirtModule(M, LegacyAARGetter(*this), OREGetter, LookupDomTree, 660 ExportSummary, ImportSummary) 661 .run(); 662 } 663 664 void getAnalysisUsage(AnalysisUsage &AU) const override { 665 AU.addRequired<AssumptionCacheTracker>(); 666 AU.addRequired<TargetLibraryInfoWrapperPass>(); 667 AU.addRequired<DominatorTreeWrapperPass>(); 668 } 669 }; 670 671 } // end anonymous namespace 672 673 INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt", 674 "Whole program devirtualization", false, false) 675 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 676 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 677 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 678 INITIALIZE_PASS_END(WholeProgramDevirt, "wholeprogramdevirt", 679 "Whole program devirtualization", false, false) 680 char WholeProgramDevirt::ID = 0; 681 682 ModulePass * 683 llvm::createWholeProgramDevirtPass(ModuleSummaryIndex *ExportSummary, 684 const ModuleSummaryIndex *ImportSummary) { 685 return new WholeProgramDevirt(ExportSummary, ImportSummary); 686 } 687 688 PreservedAnalyses WholeProgramDevirtPass::run(Module &M, 689 ModuleAnalysisManager &AM) { 690 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 691 auto AARGetter = [&](Function &F) -> AAResults & { 692 return FAM.getResult<AAManager>(F); 693 }; 694 auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & { 695 return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F); 696 }; 697 auto LookupDomTree = [&FAM](Function &F) -> DominatorTree & { 698 return FAM.getResult<DominatorTreeAnalysis>(F); 699 }; 700 if (!DevirtModule(M, AARGetter, OREGetter, LookupDomTree, ExportSummary, 701 ImportSummary) 702 .run()) 703 return PreservedAnalyses::all(); 704 return PreservedAnalyses::none(); 705 } 706 707 namespace llvm { 708 void runWholeProgramDevirtOnIndex( 709 ModuleSummaryIndex &Summary, std::set<GlobalValue::GUID> &ExportedGUIDs, 710 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 711 DevirtIndex(Summary, ExportedGUIDs, LocalWPDTargetsMap).run(); 712 } 713 714 void updateIndexWPDForExports( 715 ModuleSummaryIndex &Summary, 716 StringMap<FunctionImporter::ExportSetTy> &ExportLists, 717 std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) { 718 for (auto &T : LocalWPDTargetsMap) { 719 auto &VI = T.first; 720 // This was enforced earlier during trySingleImplDevirt. 721 assert(VI.getSummaryList().size() == 1 && 722 "Devirt of local target has more than one copy"); 723 auto &S = VI.getSummaryList()[0]; 724 const auto &ExportList = ExportLists.find(S->modulePath()); 725 if (ExportList == ExportLists.end() || 726 !ExportList->second.count(VI.getGUID())) 727 continue; 728 729 // It's been exported by a cross module import. 730 for (auto &SlotSummary : T.second) { 731 auto *TIdSum = Summary.getTypeIdSummary(SlotSummary.TypeID); 732 assert(TIdSum); 733 auto WPDRes = TIdSum->WPDRes.find(SlotSummary.ByteOffset); 734 assert(WPDRes != TIdSum->WPDRes.end()); 735 WPDRes->second.SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 736 WPDRes->second.SingleImplName, 737 Summary.getModuleHash(S->modulePath())); 738 } 739 } 740 } 741 742 } // end namespace llvm 743 744 bool DevirtModule::runForTesting( 745 Module &M, function_ref<AAResults &(Function &)> AARGetter, 746 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter, 747 function_ref<DominatorTree &(Function &)> LookupDomTree) { 748 ModuleSummaryIndex Summary(/*HaveGVs=*/false); 749 750 // Handle the command-line summary arguments. This code is for testing 751 // purposes only, so we handle errors directly. 752 if (!ClReadSummary.empty()) { 753 ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary + 754 ": "); 755 auto ReadSummaryFile = 756 ExitOnErr(errorOrToExpected(MemoryBuffer::getFile(ClReadSummary))); 757 758 yaml::Input In(ReadSummaryFile->getBuffer()); 759 In >> Summary; 760 ExitOnErr(errorCodeToError(In.error())); 761 } 762 763 bool Changed = 764 DevirtModule( 765 M, AARGetter, OREGetter, LookupDomTree, 766 ClSummaryAction == PassSummaryAction::Export ? &Summary : nullptr, 767 ClSummaryAction == PassSummaryAction::Import ? &Summary : nullptr) 768 .run(); 769 770 if (!ClWriteSummary.empty()) { 771 ExitOnError ExitOnErr( 772 "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": "); 773 std::error_code EC; 774 raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_Text); 775 ExitOnErr(errorCodeToError(EC)); 776 777 yaml::Output Out(OS); 778 Out << Summary; 779 } 780 781 return Changed; 782 } 783 784 void DevirtModule::buildTypeIdentifierMap( 785 std::vector<VTableBits> &Bits, 786 DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) { 787 DenseMap<GlobalVariable *, VTableBits *> GVToBits; 788 Bits.reserve(M.getGlobalList().size()); 789 SmallVector<MDNode *, 2> Types; 790 for (GlobalVariable &GV : M.globals()) { 791 Types.clear(); 792 GV.getMetadata(LLVMContext::MD_type, Types); 793 if (GV.isDeclaration() || Types.empty()) 794 continue; 795 796 VTableBits *&BitsPtr = GVToBits[&GV]; 797 if (!BitsPtr) { 798 Bits.emplace_back(); 799 Bits.back().GV = &GV; 800 Bits.back().ObjectSize = 801 M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType()); 802 BitsPtr = &Bits.back(); 803 } 804 805 for (MDNode *Type : Types) { 806 auto TypeID = Type->getOperand(1).get(); 807 808 uint64_t Offset = 809 cast<ConstantInt>( 810 cast<ConstantAsMetadata>(Type->getOperand(0))->getValue()) 811 ->getZExtValue(); 812 813 TypeIdMap[TypeID].insert({BitsPtr, Offset}); 814 } 815 } 816 } 817 818 Constant *DevirtModule::getPointerAtOffset(Constant *I, uint64_t Offset) { 819 if (I->getType()->isPointerTy()) { 820 if (Offset == 0) 821 return I; 822 return nullptr; 823 } 824 825 const DataLayout &DL = M.getDataLayout(); 826 827 if (auto *C = dyn_cast<ConstantStruct>(I)) { 828 const StructLayout *SL = DL.getStructLayout(C->getType()); 829 if (Offset >= SL->getSizeInBytes()) 830 return nullptr; 831 832 unsigned Op = SL->getElementContainingOffset(Offset); 833 return getPointerAtOffset(cast<Constant>(I->getOperand(Op)), 834 Offset - SL->getElementOffset(Op)); 835 } 836 if (auto *C = dyn_cast<ConstantArray>(I)) { 837 ArrayType *VTableTy = C->getType(); 838 uint64_t ElemSize = DL.getTypeAllocSize(VTableTy->getElementType()); 839 840 unsigned Op = Offset / ElemSize; 841 if (Op >= C->getNumOperands()) 842 return nullptr; 843 844 return getPointerAtOffset(cast<Constant>(I->getOperand(Op)), 845 Offset % ElemSize); 846 } 847 return nullptr; 848 } 849 850 bool DevirtModule::tryFindVirtualCallTargets( 851 std::vector<VirtualCallTarget> &TargetsForSlot, 852 const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset) { 853 for (const TypeMemberInfo &TM : TypeMemberInfos) { 854 if (!TM.Bits->GV->isConstant()) 855 return false; 856 857 Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(), 858 TM.Offset + ByteOffset); 859 if (!Ptr) 860 return false; 861 862 auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts()); 863 if (!Fn) 864 return false; 865 866 // We can disregard __cxa_pure_virtual as a possible call target, as 867 // calls to pure virtuals are UB. 868 if (Fn->getName() == "__cxa_pure_virtual") 869 continue; 870 871 TargetsForSlot.push_back({Fn, &TM}); 872 } 873 874 // Give up if we couldn't find any targets. 875 return !TargetsForSlot.empty(); 876 } 877 878 bool DevirtIndex::tryFindVirtualCallTargets( 879 std::vector<ValueInfo> &TargetsForSlot, const TypeIdCompatibleVtableInfo TIdInfo, 880 uint64_t ByteOffset) { 881 for (const TypeIdOffsetVtableInfo P : TIdInfo) { 882 // VTable initializer should have only one summary, or all copies must be 883 // linkonce/weak ODR. 884 assert(P.VTableVI.getSummaryList().size() == 1 || 885 llvm::all_of( 886 P.VTableVI.getSummaryList(), 887 [&](const std::unique_ptr<GlobalValueSummary> &Summary) { 888 return GlobalValue::isLinkOnceODRLinkage(Summary->linkage()) || 889 GlobalValue::isWeakODRLinkage(Summary->linkage()); 890 })); 891 const auto *VS = cast<GlobalVarSummary>(P.VTableVI.getSummaryList()[0].get()); 892 if (!P.VTableVI.getSummaryList()[0]->isLive()) 893 continue; 894 for (auto VTP : VS->vTableFuncs()) { 895 if (VTP.VTableOffset != P.AddressPointOffset + ByteOffset) 896 continue; 897 898 TargetsForSlot.push_back(VTP.FuncVI); 899 } 900 } 901 902 // Give up if we couldn't find any targets. 903 return !TargetsForSlot.empty(); 904 } 905 906 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo, 907 Constant *TheFn, bool &IsExported) { 908 auto Apply = [&](CallSiteInfo &CSInfo) { 909 for (auto &&VCallSite : CSInfo.CallSites) { 910 if (RemarksEnabled) 911 VCallSite.emitRemark("single-impl", 912 TheFn->stripPointerCasts()->getName(), OREGetter); 913 VCallSite.CS.setCalledFunction(ConstantExpr::getBitCast( 914 TheFn, VCallSite.CS.getCalledValue()->getType())); 915 // This use is no longer unsafe. 916 if (VCallSite.NumUnsafeUses) 917 --*VCallSite.NumUnsafeUses; 918 } 919 if (CSInfo.isExported()) 920 IsExported = true; 921 CSInfo.markDevirt(); 922 }; 923 Apply(SlotInfo.CSInfo); 924 for (auto &P : SlotInfo.ConstCSInfo) 925 Apply(P.second); 926 } 927 928 bool DevirtModule::trySingleImplDevirt( 929 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 930 VTableSlotInfo &SlotInfo, WholeProgramDevirtResolution *Res) { 931 // See if the program contains a single implementation of this virtual 932 // function. 933 Function *TheFn = TargetsForSlot[0].Fn; 934 for (auto &&Target : TargetsForSlot) 935 if (TheFn != Target.Fn) 936 return false; 937 938 // If so, update each call site to call that implementation directly. 939 if (RemarksEnabled) 940 TargetsForSlot[0].WasDevirt = true; 941 942 bool IsExported = false; 943 applySingleImplDevirt(SlotInfo, TheFn, IsExported); 944 if (!IsExported) 945 return false; 946 947 // If the only implementation has local linkage, we must promote to external 948 // to make it visible to thin LTO objects. We can only get here during the 949 // ThinLTO export phase. 950 if (TheFn->hasLocalLinkage()) { 951 std::string NewName = (TheFn->getName() + "$merged").str(); 952 953 // Since we are renaming the function, any comdats with the same name must 954 // also be renamed. This is required when targeting COFF, as the comdat name 955 // must match one of the names of the symbols in the comdat. 956 if (Comdat *C = TheFn->getComdat()) { 957 if (C->getName() == TheFn->getName()) { 958 Comdat *NewC = M.getOrInsertComdat(NewName); 959 NewC->setSelectionKind(C->getSelectionKind()); 960 for (GlobalObject &GO : M.global_objects()) 961 if (GO.getComdat() == C) 962 GO.setComdat(NewC); 963 } 964 } 965 966 TheFn->setLinkage(GlobalValue::ExternalLinkage); 967 TheFn->setVisibility(GlobalValue::HiddenVisibility); 968 TheFn->setName(NewName); 969 } 970 971 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 972 Res->SingleImplName = TheFn->getName(); 973 974 return true; 975 } 976 977 bool DevirtIndex::trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot, 978 VTableSlotSummary &SlotSummary, 979 VTableSlotInfo &SlotInfo, 980 WholeProgramDevirtResolution *Res, 981 std::set<ValueInfo> &DevirtTargets) { 982 // See if the program contains a single implementation of this virtual 983 // function. 984 auto TheFn = TargetsForSlot[0]; 985 for (auto &&Target : TargetsForSlot) 986 if (TheFn != Target) 987 return false; 988 989 // Don't devirtualize if we don't have target definition. 990 auto Size = TheFn.getSummaryList().size(); 991 if (!Size) 992 return false; 993 994 // If the summary list contains multiple summaries where at least one is 995 // a local, give up, as we won't know which (possibly promoted) name to use. 996 for (auto &S : TheFn.getSummaryList()) 997 if (GlobalValue::isLocalLinkage(S->linkage()) && Size > 1) 998 return false; 999 1000 // Collect functions devirtualized at least for one call site for stats. 1001 if (PrintSummaryDevirt) 1002 DevirtTargets.insert(TheFn); 1003 1004 auto &S = TheFn.getSummaryList()[0]; 1005 bool IsExported = false; 1006 1007 // Insert calls into the summary index so that the devirtualized targets 1008 // are eligible for import. 1009 // FIXME: Annotate type tests with hotness. For now, mark these as hot 1010 // to better ensure we have the opportunity to inline them. 1011 CalleeInfo CI(CalleeInfo::HotnessType::Hot, /* RelBF = */ 0); 1012 auto AddCalls = [&](CallSiteInfo &CSInfo) { 1013 for (auto *FS : CSInfo.SummaryTypeCheckedLoadUsers) { 1014 FS->addCall({TheFn, CI}); 1015 IsExported |= S->modulePath() != FS->modulePath(); 1016 } 1017 for (auto *FS : CSInfo.SummaryTypeTestAssumeUsers) { 1018 FS->addCall({TheFn, CI}); 1019 IsExported |= S->modulePath() != FS->modulePath(); 1020 } 1021 }; 1022 AddCalls(SlotInfo.CSInfo); 1023 for (auto &P : SlotInfo.ConstCSInfo) 1024 AddCalls(P.second); 1025 1026 if (IsExported) 1027 ExportedGUIDs.insert(TheFn.getGUID()); 1028 1029 // Record in summary for use in devirtualization during the ThinLTO import 1030 // step. 1031 Res->TheKind = WholeProgramDevirtResolution::SingleImpl; 1032 if (GlobalValue::isLocalLinkage(S->linkage())) { 1033 if (IsExported) 1034 // If target is a local function and we are exporting it by 1035 // devirtualizing a call in another module, we need to record the 1036 // promoted name. 1037 Res->SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal( 1038 TheFn.name(), ExportSummary.getModuleHash(S->modulePath())); 1039 else { 1040 LocalWPDTargetsMap[TheFn].push_back(SlotSummary); 1041 Res->SingleImplName = TheFn.name(); 1042 } 1043 } else 1044 Res->SingleImplName = TheFn.name(); 1045 1046 // Name will be empty if this thin link driven off of serialized combined 1047 // index (e.g. llvm-lto). However, WPD is not supported/invoked for the 1048 // legacy LTO API anyway. 1049 assert(!Res->SingleImplName.empty()); 1050 1051 return true; 1052 } 1053 1054 void DevirtModule::tryICallBranchFunnel( 1055 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1056 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1057 Triple T(M.getTargetTriple()); 1058 if (T.getArch() != Triple::x86_64) 1059 return; 1060 1061 if (TargetsForSlot.size() > ClThreshold) 1062 return; 1063 1064 bool HasNonDevirt = !SlotInfo.CSInfo.AllCallSitesDevirted; 1065 if (!HasNonDevirt) 1066 for (auto &P : SlotInfo.ConstCSInfo) 1067 if (!P.second.AllCallSitesDevirted) { 1068 HasNonDevirt = true; 1069 break; 1070 } 1071 1072 if (!HasNonDevirt) 1073 return; 1074 1075 FunctionType *FT = 1076 FunctionType::get(Type::getVoidTy(M.getContext()), {Int8PtrTy}, true); 1077 Function *JT; 1078 if (isa<MDString>(Slot.TypeID)) { 1079 JT = Function::Create(FT, Function::ExternalLinkage, 1080 M.getDataLayout().getProgramAddressSpace(), 1081 getGlobalName(Slot, {}, "branch_funnel"), &M); 1082 JT->setVisibility(GlobalValue::HiddenVisibility); 1083 } else { 1084 JT = Function::Create(FT, Function::InternalLinkage, 1085 M.getDataLayout().getProgramAddressSpace(), 1086 "branch_funnel", &M); 1087 } 1088 JT->addAttribute(1, Attribute::Nest); 1089 1090 std::vector<Value *> JTArgs; 1091 JTArgs.push_back(JT->arg_begin()); 1092 for (auto &T : TargetsForSlot) { 1093 JTArgs.push_back(getMemberAddr(T.TM)); 1094 JTArgs.push_back(T.Fn); 1095 } 1096 1097 BasicBlock *BB = BasicBlock::Create(M.getContext(), "", JT, nullptr); 1098 Function *Intr = 1099 Intrinsic::getDeclaration(&M, llvm::Intrinsic::icall_branch_funnel, {}); 1100 1101 auto *CI = CallInst::Create(Intr, JTArgs, "", BB); 1102 CI->setTailCallKind(CallInst::TCK_MustTail); 1103 ReturnInst::Create(M.getContext(), nullptr, BB); 1104 1105 bool IsExported = false; 1106 applyICallBranchFunnel(SlotInfo, JT, IsExported); 1107 if (IsExported) 1108 Res->TheKind = WholeProgramDevirtResolution::BranchFunnel; 1109 } 1110 1111 void DevirtModule::applyICallBranchFunnel(VTableSlotInfo &SlotInfo, 1112 Constant *JT, bool &IsExported) { 1113 auto Apply = [&](CallSiteInfo &CSInfo) { 1114 if (CSInfo.isExported()) 1115 IsExported = true; 1116 if (CSInfo.AllCallSitesDevirted) 1117 return; 1118 for (auto &&VCallSite : CSInfo.CallSites) { 1119 CallSite CS = VCallSite.CS; 1120 1121 // Jump tables are only profitable if the retpoline mitigation is enabled. 1122 Attribute FSAttr = CS.getCaller()->getFnAttribute("target-features"); 1123 if (FSAttr.hasAttribute(Attribute::None) || 1124 !FSAttr.getValueAsString().contains("+retpoline")) 1125 continue; 1126 1127 if (RemarksEnabled) 1128 VCallSite.emitRemark("branch-funnel", 1129 JT->stripPointerCasts()->getName(), OREGetter); 1130 1131 // Pass the address of the vtable in the nest register, which is r10 on 1132 // x86_64. 1133 std::vector<Type *> NewArgs; 1134 NewArgs.push_back(Int8PtrTy); 1135 for (Type *T : CS.getFunctionType()->params()) 1136 NewArgs.push_back(T); 1137 FunctionType *NewFT = 1138 FunctionType::get(CS.getFunctionType()->getReturnType(), NewArgs, 1139 CS.getFunctionType()->isVarArg()); 1140 PointerType *NewFTPtr = PointerType::getUnqual(NewFT); 1141 1142 IRBuilder<> IRB(CS.getInstruction()); 1143 std::vector<Value *> Args; 1144 Args.push_back(IRB.CreateBitCast(VCallSite.VTable, Int8PtrTy)); 1145 for (unsigned I = 0; I != CS.getNumArgOperands(); ++I) 1146 Args.push_back(CS.getArgOperand(I)); 1147 1148 CallSite NewCS; 1149 if (CS.isCall()) 1150 NewCS = IRB.CreateCall(NewFT, IRB.CreateBitCast(JT, NewFTPtr), Args); 1151 else 1152 NewCS = IRB.CreateInvoke( 1153 NewFT, IRB.CreateBitCast(JT, NewFTPtr), 1154 cast<InvokeInst>(CS.getInstruction())->getNormalDest(), 1155 cast<InvokeInst>(CS.getInstruction())->getUnwindDest(), Args); 1156 NewCS.setCallingConv(CS.getCallingConv()); 1157 1158 AttributeList Attrs = CS.getAttributes(); 1159 std::vector<AttributeSet> NewArgAttrs; 1160 NewArgAttrs.push_back(AttributeSet::get( 1161 M.getContext(), ArrayRef<Attribute>{Attribute::get( 1162 M.getContext(), Attribute::Nest)})); 1163 for (unsigned I = 0; I + 2 < Attrs.getNumAttrSets(); ++I) 1164 NewArgAttrs.push_back(Attrs.getParamAttributes(I)); 1165 NewCS.setAttributes( 1166 AttributeList::get(M.getContext(), Attrs.getFnAttributes(), 1167 Attrs.getRetAttributes(), NewArgAttrs)); 1168 1169 CS->replaceAllUsesWith(NewCS.getInstruction()); 1170 CS->eraseFromParent(); 1171 1172 // This use is no longer unsafe. 1173 if (VCallSite.NumUnsafeUses) 1174 --*VCallSite.NumUnsafeUses; 1175 } 1176 // Don't mark as devirtualized because there may be callers compiled without 1177 // retpoline mitigation, which would mean that they are lowered to 1178 // llvm.type.test and therefore require an llvm.type.test resolution for the 1179 // type identifier. 1180 }; 1181 Apply(SlotInfo.CSInfo); 1182 for (auto &P : SlotInfo.ConstCSInfo) 1183 Apply(P.second); 1184 } 1185 1186 bool DevirtModule::tryEvaluateFunctionsWithArgs( 1187 MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1188 ArrayRef<uint64_t> Args) { 1189 // Evaluate each function and store the result in each target's RetVal 1190 // field. 1191 for (VirtualCallTarget &Target : TargetsForSlot) { 1192 if (Target.Fn->arg_size() != Args.size() + 1) 1193 return false; 1194 1195 Evaluator Eval(M.getDataLayout(), nullptr); 1196 SmallVector<Constant *, 2> EvalArgs; 1197 EvalArgs.push_back( 1198 Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0))); 1199 for (unsigned I = 0; I != Args.size(); ++I) { 1200 auto *ArgTy = dyn_cast<IntegerType>( 1201 Target.Fn->getFunctionType()->getParamType(I + 1)); 1202 if (!ArgTy) 1203 return false; 1204 EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I])); 1205 } 1206 1207 Constant *RetVal; 1208 if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) || 1209 !isa<ConstantInt>(RetVal)) 1210 return false; 1211 Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue(); 1212 } 1213 return true; 1214 } 1215 1216 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1217 uint64_t TheRetVal) { 1218 for (auto Call : CSInfo.CallSites) 1219 Call.replaceAndErase( 1220 "uniform-ret-val", FnName, RemarksEnabled, OREGetter, 1221 ConstantInt::get(cast<IntegerType>(Call.CS.getType()), TheRetVal)); 1222 CSInfo.markDevirt(); 1223 } 1224 1225 bool DevirtModule::tryUniformRetValOpt( 1226 MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo, 1227 WholeProgramDevirtResolution::ByArg *Res) { 1228 // Uniform return value optimization. If all functions return the same 1229 // constant, replace all calls with that constant. 1230 uint64_t TheRetVal = TargetsForSlot[0].RetVal; 1231 for (const VirtualCallTarget &Target : TargetsForSlot) 1232 if (Target.RetVal != TheRetVal) 1233 return false; 1234 1235 if (CSInfo.isExported()) { 1236 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniformRetVal; 1237 Res->Info = TheRetVal; 1238 } 1239 1240 applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal); 1241 if (RemarksEnabled) 1242 for (auto &&Target : TargetsForSlot) 1243 Target.WasDevirt = true; 1244 return true; 1245 } 1246 1247 std::string DevirtModule::getGlobalName(VTableSlot Slot, 1248 ArrayRef<uint64_t> Args, 1249 StringRef Name) { 1250 std::string FullName = "__typeid_"; 1251 raw_string_ostream OS(FullName); 1252 OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset; 1253 for (uint64_t Arg : Args) 1254 OS << '_' << Arg; 1255 OS << '_' << Name; 1256 return OS.str(); 1257 } 1258 1259 bool DevirtModule::shouldExportConstantsAsAbsoluteSymbols() { 1260 Triple T(M.getTargetTriple()); 1261 return (T.getArch() == Triple::x86 || T.getArch() == Triple::x86_64) && 1262 T.getObjectFormat() == Triple::ELF; 1263 } 1264 1265 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1266 StringRef Name, Constant *C) { 1267 GlobalAlias *GA = GlobalAlias::create(Int8Ty, 0, GlobalValue::ExternalLinkage, 1268 getGlobalName(Slot, Args, Name), C, &M); 1269 GA->setVisibility(GlobalValue::HiddenVisibility); 1270 } 1271 1272 void DevirtModule::exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1273 StringRef Name, uint32_t Const, 1274 uint32_t &Storage) { 1275 if (shouldExportConstantsAsAbsoluteSymbols()) { 1276 exportGlobal( 1277 Slot, Args, Name, 1278 ConstantExpr::getIntToPtr(ConstantInt::get(Int32Ty, Const), Int8PtrTy)); 1279 return; 1280 } 1281 1282 Storage = Const; 1283 } 1284 1285 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, 1286 StringRef Name) { 1287 Constant *C = M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Ty); 1288 auto *GV = dyn_cast<GlobalVariable>(C); 1289 if (GV) 1290 GV->setVisibility(GlobalValue::HiddenVisibility); 1291 return C; 1292 } 1293 1294 Constant *DevirtModule::importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, 1295 StringRef Name, IntegerType *IntTy, 1296 uint32_t Storage) { 1297 if (!shouldExportConstantsAsAbsoluteSymbols()) 1298 return ConstantInt::get(IntTy, Storage); 1299 1300 Constant *C = importGlobal(Slot, Args, Name); 1301 auto *GV = cast<GlobalVariable>(C->stripPointerCasts()); 1302 C = ConstantExpr::getPtrToInt(C, IntTy); 1303 1304 // We only need to set metadata if the global is newly created, in which 1305 // case it would not have hidden visibility. 1306 if (GV->hasMetadata(LLVMContext::MD_absolute_symbol)) 1307 return C; 1308 1309 auto SetAbsRange = [&](uint64_t Min, uint64_t Max) { 1310 auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min)); 1311 auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max)); 1312 GV->setMetadata(LLVMContext::MD_absolute_symbol, 1313 MDNode::get(M.getContext(), {MinC, MaxC})); 1314 }; 1315 unsigned AbsWidth = IntTy->getBitWidth(); 1316 if (AbsWidth == IntPtrTy->getBitWidth()) 1317 SetAbsRange(~0ull, ~0ull); // Full set. 1318 else 1319 SetAbsRange(0, 1ull << AbsWidth); 1320 return C; 1321 } 1322 1323 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, 1324 bool IsOne, 1325 Constant *UniqueMemberAddr) { 1326 for (auto &&Call : CSInfo.CallSites) { 1327 IRBuilder<> B(Call.CS.getInstruction()); 1328 Value *Cmp = 1329 B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE, 1330 B.CreateBitCast(Call.VTable, Int8PtrTy), UniqueMemberAddr); 1331 Cmp = B.CreateZExt(Cmp, Call.CS->getType()); 1332 Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter, 1333 Cmp); 1334 } 1335 CSInfo.markDevirt(); 1336 } 1337 1338 Constant *DevirtModule::getMemberAddr(const TypeMemberInfo *M) { 1339 Constant *C = ConstantExpr::getBitCast(M->Bits->GV, Int8PtrTy); 1340 return ConstantExpr::getGetElementPtr(Int8Ty, C, 1341 ConstantInt::get(Int64Ty, M->Offset)); 1342 } 1343 1344 bool DevirtModule::tryUniqueRetValOpt( 1345 unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot, 1346 CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res, 1347 VTableSlot Slot, ArrayRef<uint64_t> Args) { 1348 // IsOne controls whether we look for a 0 or a 1. 1349 auto tryUniqueRetValOptFor = [&](bool IsOne) { 1350 const TypeMemberInfo *UniqueMember = nullptr; 1351 for (const VirtualCallTarget &Target : TargetsForSlot) { 1352 if (Target.RetVal == (IsOne ? 1 : 0)) { 1353 if (UniqueMember) 1354 return false; 1355 UniqueMember = Target.TM; 1356 } 1357 } 1358 1359 // We should have found a unique member or bailed out by now. We already 1360 // checked for a uniform return value in tryUniformRetValOpt. 1361 assert(UniqueMember); 1362 1363 Constant *UniqueMemberAddr = getMemberAddr(UniqueMember); 1364 if (CSInfo.isExported()) { 1365 Res->TheKind = WholeProgramDevirtResolution::ByArg::UniqueRetVal; 1366 Res->Info = IsOne; 1367 1368 exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr); 1369 } 1370 1371 // Replace each call with the comparison. 1372 applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne, 1373 UniqueMemberAddr); 1374 1375 // Update devirtualization statistics for targets. 1376 if (RemarksEnabled) 1377 for (auto &&Target : TargetsForSlot) 1378 Target.WasDevirt = true; 1379 1380 return true; 1381 }; 1382 1383 if (BitWidth == 1) { 1384 if (tryUniqueRetValOptFor(true)) 1385 return true; 1386 if (tryUniqueRetValOptFor(false)) 1387 return true; 1388 } 1389 return false; 1390 } 1391 1392 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName, 1393 Constant *Byte, Constant *Bit) { 1394 for (auto Call : CSInfo.CallSites) { 1395 auto *RetType = cast<IntegerType>(Call.CS.getType()); 1396 IRBuilder<> B(Call.CS.getInstruction()); 1397 Value *Addr = 1398 B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte); 1399 if (RetType->getBitWidth() == 1) { 1400 Value *Bits = B.CreateLoad(Int8Ty, Addr); 1401 Value *BitsAndBit = B.CreateAnd(Bits, Bit); 1402 auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0)); 1403 Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled, 1404 OREGetter, IsBitSet); 1405 } else { 1406 Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo()); 1407 Value *Val = B.CreateLoad(RetType, ValAddr); 1408 Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled, 1409 OREGetter, Val); 1410 } 1411 } 1412 CSInfo.markDevirt(); 1413 } 1414 1415 bool DevirtModule::tryVirtualConstProp( 1416 MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo, 1417 WholeProgramDevirtResolution *Res, VTableSlot Slot) { 1418 // This only works if the function returns an integer. 1419 auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType()); 1420 if (!RetType) 1421 return false; 1422 unsigned BitWidth = RetType->getBitWidth(); 1423 if (BitWidth > 64) 1424 return false; 1425 1426 // Make sure that each function is defined, does not access memory, takes at 1427 // least one argument, does not use its first argument (which we assume is 1428 // 'this'), and has the same return type. 1429 // 1430 // Note that we test whether this copy of the function is readnone, rather 1431 // than testing function attributes, which must hold for any copy of the 1432 // function, even a less optimized version substituted at link time. This is 1433 // sound because the virtual constant propagation optimizations effectively 1434 // inline all implementations of the virtual function into each call site, 1435 // rather than using function attributes to perform local optimization. 1436 for (VirtualCallTarget &Target : TargetsForSlot) { 1437 if (Target.Fn->isDeclaration() || 1438 computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) != 1439 MAK_ReadNone || 1440 Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() || 1441 Target.Fn->getReturnType() != RetType) 1442 return false; 1443 } 1444 1445 for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) { 1446 if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first)) 1447 continue; 1448 1449 WholeProgramDevirtResolution::ByArg *ResByArg = nullptr; 1450 if (Res) 1451 ResByArg = &Res->ResByArg[CSByConstantArg.first]; 1452 1453 if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg)) 1454 continue; 1455 1456 if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second, 1457 ResByArg, Slot, CSByConstantArg.first)) 1458 continue; 1459 1460 // Find an allocation offset in bits in all vtables associated with the 1461 // type. 1462 uint64_t AllocBefore = 1463 findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth); 1464 uint64_t AllocAfter = 1465 findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth); 1466 1467 // Calculate the total amount of padding needed to store a value at both 1468 // ends of the object. 1469 uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0; 1470 for (auto &&Target : TargetsForSlot) { 1471 TotalPaddingBefore += std::max<int64_t>( 1472 (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0); 1473 TotalPaddingAfter += std::max<int64_t>( 1474 (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0); 1475 } 1476 1477 // If the amount of padding is too large, give up. 1478 // FIXME: do something smarter here. 1479 if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128) 1480 continue; 1481 1482 // Calculate the offset to the value as a (possibly negative) byte offset 1483 // and (if applicable) a bit offset, and store the values in the targets. 1484 int64_t OffsetByte; 1485 uint64_t OffsetBit; 1486 if (TotalPaddingBefore <= TotalPaddingAfter) 1487 setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte, 1488 OffsetBit); 1489 else 1490 setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte, 1491 OffsetBit); 1492 1493 if (RemarksEnabled) 1494 for (auto &&Target : TargetsForSlot) 1495 Target.WasDevirt = true; 1496 1497 1498 if (CSByConstantArg.second.isExported()) { 1499 ResByArg->TheKind = WholeProgramDevirtResolution::ByArg::VirtualConstProp; 1500 exportConstant(Slot, CSByConstantArg.first, "byte", OffsetByte, 1501 ResByArg->Byte); 1502 exportConstant(Slot, CSByConstantArg.first, "bit", 1ULL << OffsetBit, 1503 ResByArg->Bit); 1504 } 1505 1506 // Rewrite each call to a load from OffsetByte/OffsetBit. 1507 Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte); 1508 Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit); 1509 applyVirtualConstProp(CSByConstantArg.second, 1510 TargetsForSlot[0].Fn->getName(), ByteConst, BitConst); 1511 } 1512 return true; 1513 } 1514 1515 void DevirtModule::rebuildGlobal(VTableBits &B) { 1516 if (B.Before.Bytes.empty() && B.After.Bytes.empty()) 1517 return; 1518 1519 // Align the before byte array to the global's minimum alignment so that we 1520 // don't break any alignment requirements on the global. 1521 unsigned Align = B.GV->getAlignment(); 1522 if (Align == 0) 1523 Align = M.getDataLayout().getABITypeAlignment(B.GV->getValueType()); 1524 B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), Align)); 1525 1526 // Before was stored in reverse order; flip it now. 1527 for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I) 1528 std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]); 1529 1530 // Build an anonymous global containing the before bytes, followed by the 1531 // original initializer, followed by the after bytes. 1532 auto NewInit = ConstantStruct::getAnon( 1533 {ConstantDataArray::get(M.getContext(), B.Before.Bytes), 1534 B.GV->getInitializer(), 1535 ConstantDataArray::get(M.getContext(), B.After.Bytes)}); 1536 auto NewGV = 1537 new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(), 1538 GlobalVariable::PrivateLinkage, NewInit, "", B.GV); 1539 NewGV->setSection(B.GV->getSection()); 1540 NewGV->setComdat(B.GV->getComdat()); 1541 NewGV->setAlignment(B.GV->getAlignment()); 1542 1543 // Copy the original vtable's metadata to the anonymous global, adjusting 1544 // offsets as required. 1545 NewGV->copyMetadata(B.GV, B.Before.Bytes.size()); 1546 1547 // Build an alias named after the original global, pointing at the second 1548 // element (the original initializer). 1549 auto Alias = GlobalAlias::create( 1550 B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "", 1551 ConstantExpr::getGetElementPtr( 1552 NewInit->getType(), NewGV, 1553 ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0), 1554 ConstantInt::get(Int32Ty, 1)}), 1555 &M); 1556 Alias->setVisibility(B.GV->getVisibility()); 1557 Alias->takeName(B.GV); 1558 1559 B.GV->replaceAllUsesWith(Alias); 1560 B.GV->eraseFromParent(); 1561 } 1562 1563 bool DevirtModule::areRemarksEnabled() { 1564 const auto &FL = M.getFunctionList(); 1565 for (const Function &Fn : FL) { 1566 const auto &BBL = Fn.getBasicBlockList(); 1567 if (BBL.empty()) 1568 continue; 1569 auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front()); 1570 return DI.isEnabled(); 1571 } 1572 return false; 1573 } 1574 1575 void DevirtModule::scanTypeTestUsers(Function *TypeTestFunc, 1576 Function *AssumeFunc) { 1577 // Find all virtual calls via a virtual table pointer %p under an assumption 1578 // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p 1579 // points to a member of the type identifier %md. Group calls by (type ID, 1580 // offset) pair (effectively the identity of the virtual function) and store 1581 // to CallSlots. 1582 DenseSet<CallSite> SeenCallSites; 1583 for (auto I = TypeTestFunc->use_begin(), E = TypeTestFunc->use_end(); 1584 I != E;) { 1585 auto CI = dyn_cast<CallInst>(I->getUser()); 1586 ++I; 1587 if (!CI) 1588 continue; 1589 1590 // Search for virtual calls based on %p and add them to DevirtCalls. 1591 SmallVector<DevirtCallSite, 1> DevirtCalls; 1592 SmallVector<CallInst *, 1> Assumes; 1593 auto &DT = LookupDomTree(*CI->getFunction()); 1594 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT); 1595 1596 // If we found any, add them to CallSlots. 1597 if (!Assumes.empty()) { 1598 Metadata *TypeId = 1599 cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata(); 1600 Value *Ptr = CI->getArgOperand(0)->stripPointerCasts(); 1601 for (DevirtCallSite Call : DevirtCalls) { 1602 // Only add this CallSite if we haven't seen it before. The vtable 1603 // pointer may have been CSE'd with pointers from other call sites, 1604 // and we don't want to process call sites multiple times. We can't 1605 // just skip the vtable Ptr if it has been seen before, however, since 1606 // it may be shared by type tests that dominate different calls. 1607 if (SeenCallSites.insert(Call.CS).second) 1608 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS, nullptr); 1609 } 1610 } 1611 1612 // We no longer need the assumes or the type test. 1613 for (auto Assume : Assumes) 1614 Assume->eraseFromParent(); 1615 // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we 1616 // may use the vtable argument later. 1617 if (CI->use_empty()) 1618 CI->eraseFromParent(); 1619 } 1620 } 1621 1622 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) { 1623 Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test); 1624 1625 for (auto I = TypeCheckedLoadFunc->use_begin(), 1626 E = TypeCheckedLoadFunc->use_end(); 1627 I != E;) { 1628 auto CI = dyn_cast<CallInst>(I->getUser()); 1629 ++I; 1630 if (!CI) 1631 continue; 1632 1633 Value *Ptr = CI->getArgOperand(0); 1634 Value *Offset = CI->getArgOperand(1); 1635 Value *TypeIdValue = CI->getArgOperand(2); 1636 Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata(); 1637 1638 SmallVector<DevirtCallSite, 1> DevirtCalls; 1639 SmallVector<Instruction *, 1> LoadedPtrs; 1640 SmallVector<Instruction *, 1> Preds; 1641 bool HasNonCallUses = false; 1642 auto &DT = LookupDomTree(*CI->getFunction()); 1643 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds, 1644 HasNonCallUses, CI, DT); 1645 1646 // Start by generating "pessimistic" code that explicitly loads the function 1647 // pointer from the vtable and performs the type check. If possible, we will 1648 // eliminate the load and the type check later. 1649 1650 // If possible, only generate the load at the point where it is used. 1651 // This helps avoid unnecessary spills. 1652 IRBuilder<> LoadB( 1653 (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI); 1654 Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset); 1655 Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy)); 1656 Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr); 1657 1658 for (Instruction *LoadedPtr : LoadedPtrs) { 1659 LoadedPtr->replaceAllUsesWith(LoadedValue); 1660 LoadedPtr->eraseFromParent(); 1661 } 1662 1663 // Likewise for the type test. 1664 IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI); 1665 CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue}); 1666 1667 for (Instruction *Pred : Preds) { 1668 Pred->replaceAllUsesWith(TypeTestCall); 1669 Pred->eraseFromParent(); 1670 } 1671 1672 // We have already erased any extractvalue instructions that refer to the 1673 // intrinsic call, but the intrinsic may have other non-extractvalue uses 1674 // (although this is unlikely). In that case, explicitly build a pair and 1675 // RAUW it. 1676 if (!CI->use_empty()) { 1677 Value *Pair = UndefValue::get(CI->getType()); 1678 IRBuilder<> B(CI); 1679 Pair = B.CreateInsertValue(Pair, LoadedValue, {0}); 1680 Pair = B.CreateInsertValue(Pair, TypeTestCall, {1}); 1681 CI->replaceAllUsesWith(Pair); 1682 } 1683 1684 // The number of unsafe uses is initially the number of uses. 1685 auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall]; 1686 NumUnsafeUses = DevirtCalls.size(); 1687 1688 // If the function pointer has a non-call user, we cannot eliminate the type 1689 // check, as one of those users may eventually call the pointer. Increment 1690 // the unsafe use count to make sure it cannot reach zero. 1691 if (HasNonCallUses) 1692 ++NumUnsafeUses; 1693 for (DevirtCallSite Call : DevirtCalls) { 1694 CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS, 1695 &NumUnsafeUses); 1696 } 1697 1698 CI->eraseFromParent(); 1699 } 1700 } 1701 1702 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) { 1703 auto *TypeId = dyn_cast<MDString>(Slot.TypeID); 1704 if (!TypeId) 1705 return; 1706 const TypeIdSummary *TidSummary = 1707 ImportSummary->getTypeIdSummary(TypeId->getString()); 1708 if (!TidSummary) 1709 return; 1710 auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset); 1711 if (ResI == TidSummary->WPDRes.end()) 1712 return; 1713 const WholeProgramDevirtResolution &Res = ResI->second; 1714 1715 if (Res.TheKind == WholeProgramDevirtResolution::SingleImpl) { 1716 assert(!Res.SingleImplName.empty()); 1717 // The type of the function in the declaration is irrelevant because every 1718 // call site will cast it to the correct type. 1719 Constant *SingleImpl = 1720 cast<Constant>(M.getOrInsertFunction(Res.SingleImplName, 1721 Type::getVoidTy(M.getContext())) 1722 .getCallee()); 1723 1724 // This is the import phase so we should not be exporting anything. 1725 bool IsExported = false; 1726 applySingleImplDevirt(SlotInfo, SingleImpl, IsExported); 1727 assert(!IsExported); 1728 } 1729 1730 for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) { 1731 auto I = Res.ResByArg.find(CSByConstantArg.first); 1732 if (I == Res.ResByArg.end()) 1733 continue; 1734 auto &ResByArg = I->second; 1735 // FIXME: We should figure out what to do about the "function name" argument 1736 // to the apply* functions, as the function names are unavailable during the 1737 // importing phase. For now we just pass the empty string. This does not 1738 // impact correctness because the function names are just used for remarks. 1739 switch (ResByArg.TheKind) { 1740 case WholeProgramDevirtResolution::ByArg::UniformRetVal: 1741 applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info); 1742 break; 1743 case WholeProgramDevirtResolution::ByArg::UniqueRetVal: { 1744 Constant *UniqueMemberAddr = 1745 importGlobal(Slot, CSByConstantArg.first, "unique_member"); 1746 applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info, 1747 UniqueMemberAddr); 1748 break; 1749 } 1750 case WholeProgramDevirtResolution::ByArg::VirtualConstProp: { 1751 Constant *Byte = importConstant(Slot, CSByConstantArg.first, "byte", 1752 Int32Ty, ResByArg.Byte); 1753 Constant *Bit = importConstant(Slot, CSByConstantArg.first, "bit", Int8Ty, 1754 ResByArg.Bit); 1755 applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit); 1756 break; 1757 } 1758 default: 1759 break; 1760 } 1761 } 1762 1763 if (Res.TheKind == WholeProgramDevirtResolution::BranchFunnel) { 1764 // The type of the function is irrelevant, because it's bitcast at calls 1765 // anyhow. 1766 Constant *JT = cast<Constant>( 1767 M.getOrInsertFunction(getGlobalName(Slot, {}, "branch_funnel"), 1768 Type::getVoidTy(M.getContext())) 1769 .getCallee()); 1770 bool IsExported = false; 1771 applyICallBranchFunnel(SlotInfo, JT, IsExported); 1772 assert(!IsExported); 1773 } 1774 } 1775 1776 void DevirtModule::removeRedundantTypeTests() { 1777 auto True = ConstantInt::getTrue(M.getContext()); 1778 for (auto &&U : NumUnsafeUsesForTypeTest) { 1779 if (U.second == 0) { 1780 U.first->replaceAllUsesWith(True); 1781 U.first->eraseFromParent(); 1782 } 1783 } 1784 } 1785 1786 bool DevirtModule::run() { 1787 // If only some of the modules were split, we cannot correctly perform 1788 // this transformation. We already checked for the presense of type tests 1789 // with partially split modules during the thin link, and would have emitted 1790 // an error if any were found, so here we can simply return. 1791 if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) || 1792 (ImportSummary && ImportSummary->partiallySplitLTOUnits())) 1793 return false; 1794 1795 Function *TypeTestFunc = 1796 M.getFunction(Intrinsic::getName(Intrinsic::type_test)); 1797 Function *TypeCheckedLoadFunc = 1798 M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load)); 1799 Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume)); 1800 1801 // Normally if there are no users of the devirtualization intrinsics in the 1802 // module, this pass has nothing to do. But if we are exporting, we also need 1803 // to handle any users that appear only in the function summaries. 1804 if (!ExportSummary && 1805 (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc || 1806 AssumeFunc->use_empty()) && 1807 (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty())) 1808 return false; 1809 1810 if (TypeTestFunc && AssumeFunc) 1811 scanTypeTestUsers(TypeTestFunc, AssumeFunc); 1812 1813 if (TypeCheckedLoadFunc) 1814 scanTypeCheckedLoadUsers(TypeCheckedLoadFunc); 1815 1816 if (ImportSummary) { 1817 for (auto &S : CallSlots) 1818 importResolution(S.first, S.second); 1819 1820 removeRedundantTypeTests(); 1821 1822 // The rest of the code is only necessary when exporting or during regular 1823 // LTO, so we are done. 1824 return true; 1825 } 1826 1827 // Rebuild type metadata into a map for easy lookup. 1828 std::vector<VTableBits> Bits; 1829 DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap; 1830 buildTypeIdentifierMap(Bits, TypeIdMap); 1831 if (TypeIdMap.empty()) 1832 return true; 1833 1834 // Collect information from summary about which calls to try to devirtualize. 1835 if (ExportSummary) { 1836 DenseMap<GlobalValue::GUID, TinyPtrVector<Metadata *>> MetadataByGUID; 1837 for (auto &P : TypeIdMap) { 1838 if (auto *TypeId = dyn_cast<MDString>(P.first)) 1839 MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back( 1840 TypeId); 1841 } 1842 1843 for (auto &P : *ExportSummary) { 1844 for (auto &S : P.second.SummaryList) { 1845 auto *FS = dyn_cast<FunctionSummary>(S.get()); 1846 if (!FS) 1847 continue; 1848 // FIXME: Only add live functions. 1849 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 1850 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 1851 CallSlots[{MD, VF.Offset}] 1852 .CSInfo.markSummaryHasTypeTestAssumeUsers(); 1853 } 1854 } 1855 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 1856 for (Metadata *MD : MetadataByGUID[VF.GUID]) { 1857 CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 1858 } 1859 } 1860 for (const FunctionSummary::ConstVCall &VC : 1861 FS->type_test_assume_const_vcalls()) { 1862 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 1863 CallSlots[{MD, VC.VFunc.Offset}] 1864 .ConstCSInfo[VC.Args] 1865 .markSummaryHasTypeTestAssumeUsers(); 1866 } 1867 } 1868 for (const FunctionSummary::ConstVCall &VC : 1869 FS->type_checked_load_const_vcalls()) { 1870 for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) { 1871 CallSlots[{MD, VC.VFunc.Offset}] 1872 .ConstCSInfo[VC.Args] 1873 .addSummaryTypeCheckedLoadUser(FS); 1874 } 1875 } 1876 } 1877 } 1878 } 1879 1880 // For each (type, offset) pair: 1881 bool DidVirtualConstProp = false; 1882 std::map<std::string, Function*> DevirtTargets; 1883 for (auto &S : CallSlots) { 1884 // Search each of the members of the type identifier for the virtual 1885 // function implementation at offset S.first.ByteOffset, and add to 1886 // TargetsForSlot. 1887 std::vector<VirtualCallTarget> TargetsForSlot; 1888 if (tryFindVirtualCallTargets(TargetsForSlot, TypeIdMap[S.first.TypeID], 1889 S.first.ByteOffset)) { 1890 WholeProgramDevirtResolution *Res = nullptr; 1891 if (ExportSummary && isa<MDString>(S.first.TypeID)) 1892 Res = &ExportSummary 1893 ->getOrInsertTypeIdSummary( 1894 cast<MDString>(S.first.TypeID)->getString()) 1895 .WPDRes[S.first.ByteOffset]; 1896 1897 if (!trySingleImplDevirt(TargetsForSlot, S.second, Res)) { 1898 DidVirtualConstProp |= 1899 tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first); 1900 1901 tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first); 1902 } 1903 1904 // Collect functions devirtualized at least for one call site for stats. 1905 if (RemarksEnabled) 1906 for (const auto &T : TargetsForSlot) 1907 if (T.WasDevirt) 1908 DevirtTargets[T.Fn->getName()] = T.Fn; 1909 } 1910 1911 // CFI-specific: if we are exporting and any llvm.type.checked.load 1912 // intrinsics were *not* devirtualized, we need to add the resulting 1913 // llvm.type.test intrinsics to the function summaries so that the 1914 // LowerTypeTests pass will export them. 1915 if (ExportSummary && isa<MDString>(S.first.TypeID)) { 1916 auto GUID = 1917 GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString()); 1918 for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers) 1919 FS->addTypeTest(GUID); 1920 for (auto &CCS : S.second.ConstCSInfo) 1921 for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers) 1922 FS->addTypeTest(GUID); 1923 } 1924 } 1925 1926 if (RemarksEnabled) { 1927 // Generate remarks for each devirtualized function. 1928 for (const auto &DT : DevirtTargets) { 1929 Function *F = DT.second; 1930 1931 using namespace ore; 1932 OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F) 1933 << "devirtualized " 1934 << NV("FunctionName", DT.first)); 1935 } 1936 } 1937 1938 removeRedundantTypeTests(); 1939 1940 // Rebuild each global we touched as part of virtual constant propagation to 1941 // include the before and after bytes. 1942 if (DidVirtualConstProp) 1943 for (VTableBits &B : Bits) 1944 rebuildGlobal(B); 1945 1946 return true; 1947 } 1948 1949 void DevirtIndex::run() { 1950 if (ExportSummary.typeIdCompatibleVtableMap().empty()) 1951 return; 1952 1953 DenseMap<GlobalValue::GUID, std::vector<StringRef>> NameByGUID; 1954 for (auto &P : ExportSummary.typeIdCompatibleVtableMap()) { 1955 NameByGUID[GlobalValue::getGUID(P.first)].push_back(P.first); 1956 } 1957 1958 // Collect information from summary about which calls to try to devirtualize. 1959 for (auto &P : ExportSummary) { 1960 for (auto &S : P.second.SummaryList) { 1961 auto *FS = dyn_cast<FunctionSummary>(S.get()); 1962 if (!FS) 1963 continue; 1964 // FIXME: Only add live functions. 1965 for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) { 1966 for (StringRef Name : NameByGUID[VF.GUID]) { 1967 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS); 1968 } 1969 } 1970 for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) { 1971 for (StringRef Name : NameByGUID[VF.GUID]) { 1972 CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS); 1973 } 1974 } 1975 for (const FunctionSummary::ConstVCall &VC : 1976 FS->type_test_assume_const_vcalls()) { 1977 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 1978 CallSlots[{Name, VC.VFunc.Offset}] 1979 .ConstCSInfo[VC.Args] 1980 .addSummaryTypeTestAssumeUser(FS); 1981 } 1982 } 1983 for (const FunctionSummary::ConstVCall &VC : 1984 FS->type_checked_load_const_vcalls()) { 1985 for (StringRef Name : NameByGUID[VC.VFunc.GUID]) { 1986 CallSlots[{Name, VC.VFunc.Offset}] 1987 .ConstCSInfo[VC.Args] 1988 .addSummaryTypeCheckedLoadUser(FS); 1989 } 1990 } 1991 } 1992 } 1993 1994 std::set<ValueInfo> DevirtTargets; 1995 // For each (type, offset) pair: 1996 for (auto &S : CallSlots) { 1997 // Search each of the members of the type identifier for the virtual 1998 // function implementation at offset S.first.ByteOffset, and add to 1999 // TargetsForSlot. 2000 std::vector<ValueInfo> TargetsForSlot; 2001 auto TidSummary = ExportSummary.getTypeIdCompatibleVtableSummary(S.first.TypeID); 2002 assert(TidSummary); 2003 if (tryFindVirtualCallTargets(TargetsForSlot, *TidSummary, 2004 S.first.ByteOffset)) { 2005 WholeProgramDevirtResolution *Res = 2006 &ExportSummary.getOrInsertTypeIdSummary(S.first.TypeID) 2007 .WPDRes[S.first.ByteOffset]; 2008 2009 if (!trySingleImplDevirt(TargetsForSlot, S.first, S.second, Res, 2010 DevirtTargets)) 2011 continue; 2012 } 2013 } 2014 2015 // Optionally have the thin link print message for each devirtualized 2016 // function. 2017 if (PrintSummaryDevirt) 2018 for (const auto &DT : DevirtTargets) 2019 errs() << "Devirtualized call to " << DT << "\n"; 2020 2021 return; 2022 } 2023