LLVM  7.0.0svn
WholeProgramDevirt.cpp
Go to the documentation of this file.
1 //===- WholeProgramDevirt.cpp - Whole program virtual call optimization ---===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass implements whole program optimization of virtual calls in cases
11 // where we know (via !type metadata) that the list of callees is fixed. This
12 // includes the following:
13 // - Single implementation devirtualization: if a virtual call has a single
14 // possible callee, replace all calls with a direct call to that callee.
15 // - Virtual constant propagation: if the virtual function's return type is an
16 // integer <=64 bits and all possible callees are readnone, for each class and
17 // each list of constant arguments: evaluate the function, store the return
18 // value alongside the virtual table, and rewrite each virtual call as a load
19 // from the virtual table.
20 // - Uniform return value optimization: if the conditions for virtual constant
21 // propagation hold and each function returns the same constant value, replace
22 // each virtual call with that constant.
23 // - Unique return value optimization for i1 return values: if the conditions
24 // for virtual constant propagation hold and a single vtable's function
25 // returns 0, or a single vtable's function returns 1, replace each virtual
26 // call with a comparison of the vptr against that vtable's address.
27 //
28 // This pass is intended to be used during the regular and thin LTO pipelines.
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). During
33 // ThinLTO, the pass operates in two phases:
34 // - Export phase: this is run during the thin link over a single merged module
35 // that contains all vtables with !type metadata that participate in the link.
36 // The pass computes a resolution for each virtual call and stores it in the
37 // type identifier summary.
38 // - Import phase: this is run during the thin backends over the individual
39 // modules. The pass applies the resolutions previously computed during the
40 // import phase to each eligible virtual call.
41 //
42 //===----------------------------------------------------------------------===//
43 
45 #include "llvm/ADT/ArrayRef.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/DenseMapInfo.h"
48 #include "llvm/ADT/DenseSet.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/IR/CallSite.h"
57 #include "llvm/IR/Constants.h"
58 #include "llvm/IR/DataLayout.h"
59 #include "llvm/IR/DebugLoc.h"
60 #include "llvm/IR/DerivedTypes.h"
61 #include "llvm/IR/Function.h"
62 #include "llvm/IR/GlobalAlias.h"
63 #include "llvm/IR/GlobalVariable.h"
64 #include "llvm/IR/IRBuilder.h"
65 #include "llvm/IR/InstrTypes.h"
66 #include "llvm/IR/Instruction.h"
67 #include "llvm/IR/Instructions.h"
68 #include "llvm/IR/Intrinsics.h"
69 #include "llvm/IR/LLVMContext.h"
70 #include "llvm/IR/Metadata.h"
71 #include "llvm/IR/Module.h"
73 #include "llvm/Pass.h"
74 #include "llvm/PassRegistry.h"
75 #include "llvm/PassSupport.h"
76 #include "llvm/Support/Casting.h"
77 #include "llvm/Support/Error.h"
80 #include "llvm/Transforms/IPO.h"
83 #include <algorithm>
84 #include <cstddef>
85 #include <map>
86 #include <set>
87 #include <string>
88 
89 using namespace llvm;
90 using namespace wholeprogramdevirt;
91 
92 #define DEBUG_TYPE "wholeprogramdevirt"
93 
95  "wholeprogramdevirt-summary-action",
96  cl::desc("What to do with the summary when running this pass"),
97  cl::values(clEnumValN(PassSummaryAction::None, "none", "Do nothing"),
99  "Import typeid resolutions from summary and globals"),
101  "Export typeid resolutions to summary and globals")),
102  cl::Hidden);
103 
105  "wholeprogramdevirt-read-summary",
106  cl::desc("Read summary from given YAML file before running pass"),
107  cl::Hidden);
108 
110  "wholeprogramdevirt-write-summary",
111  cl::desc("Write summary to given YAML file after running pass"),
112  cl::Hidden);
113 
114 static cl::opt<unsigned>
115  ClThreshold("wholeprogramdevirt-branch-funnel-threshold", cl::Hidden,
117  cl::desc("Maximum number of call targets per "
118  "call site to enable branch funnels"));
119 
120 // Find the minimum offset that we may store a value of size Size bits at. If
121 // IsAfter is set, look for an offset before the object, otherwise look for an
122 // offset after the object.
123 uint64_t
125  bool IsAfter, uint64_t Size) {
126  // Find a minimum offset taking into account only vtable sizes.
127  uint64_t MinByte = 0;
128  for (const VirtualCallTarget &Target : Targets) {
129  if (IsAfter)
130  MinByte = std::max(MinByte, Target.minAfterBytes());
131  else
132  MinByte = std::max(MinByte, Target.minBeforeBytes());
133  }
134 
135  // Build a vector of arrays of bytes covering, for each target, a slice of the
136  // used region (see AccumBitVector::BytesUsed in
137  // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively,
138  // this aligns the used regions to start at MinByte.
139  //
140  // In this example, A, B and C are vtables, # is a byte already allocated for
141  // a virtual function pointer, AAAA... (etc.) are the used regions for the
142  // vtables and Offset(X) is the value computed for the Offset variable below
143  // for X.
144  //
145  // Offset(A)
146  // | |
147  // |MinByte
148  // A: ################AAAAAAAA|AAAAAAAA
149  // B: ########BBBBBBBBBBBBBBBB|BBBB
150  // C: ########################|CCCCCCCCCCCCCCCC
151  // | Offset(B) |
152  //
153  // This code produces the slices of A, B and C that appear after the divider
154  // at MinByte.
155  std::vector<ArrayRef<uint8_t>> Used;
156  for (const VirtualCallTarget &Target : Targets) {
157  ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed
158  : Target.TM->Bits->Before.BytesUsed;
159  uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes()
160  : MinByte - Target.minBeforeBytes();
161 
162  // Disregard used regions that are smaller than Offset. These are
163  // effectively all-free regions that do not need to be checked.
164  if (VTUsed.size() > Offset)
165  Used.push_back(VTUsed.slice(Offset));
166  }
167 
168  if (Size == 1) {
169  // Find a free bit in each member of Used.
170  for (unsigned I = 0;; ++I) {
171  uint8_t BitsUsed = 0;
172  for (auto &&B : Used)
173  if (I < B.size())
174  BitsUsed |= B[I];
175  if (BitsUsed != 0xff)
176  return (MinByte + I) * 8 +
177  countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined);
178  }
179  } else {
180  // Find a free (Size/8) byte region in each member of Used.
181  // FIXME: see if alignment helps.
182  for (unsigned I = 0;; ++I) {
183  for (auto &&B : Used) {
184  unsigned Byte = 0;
185  while ((I + Byte) < B.size() && Byte < (Size / 8)) {
186  if (B[I + Byte])
187  goto NextI;
188  ++Byte;
189  }
190  }
191  return (MinByte + I) * 8;
192  NextI:;
193  }
194  }
195 }
196 
198  MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore,
199  unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
200  if (BitWidth == 1)
201  OffsetByte = -(AllocBefore / 8 + 1);
202  else
203  OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8);
204  OffsetBit = AllocBefore % 8;
205 
206  for (VirtualCallTarget &Target : Targets) {
207  if (BitWidth == 1)
208  Target.setBeforeBit(AllocBefore);
209  else
210  Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8);
211  }
212 }
213 
215  MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter,
216  unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
217  if (BitWidth == 1)
218  OffsetByte = AllocAfter / 8;
219  else
220  OffsetByte = (AllocAfter + 7) / 8;
221  OffsetBit = AllocAfter % 8;
222 
223  for (VirtualCallTarget &Target : Targets) {
224  if (BitWidth == 1)
225  Target.setAfterBit(AllocAfter);
226  else
227  Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8);
228  }
229 }
230 
232  : Fn(Fn), TM(TM),
233  IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {}
234 
235 namespace {
236 
237 // A slot in a set of virtual tables. The TypeID identifies the set of virtual
238 // tables, and the ByteOffset is the offset in bytes from the address point to
239 // the virtual function pointer.
240 struct VTableSlot {
241  Metadata *TypeID;
242  uint64_t ByteOffset;
243 };
244 
245 } // end anonymous namespace
246 
247 namespace llvm {
248 
249 template <> struct DenseMapInfo<VTableSlot> {
250  static VTableSlot getEmptyKey() {
253  }
254  static VTableSlot getTombstoneKey() {
257  }
258  static unsigned getHashValue(const VTableSlot &I) {
259  return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^
261  }
262  static bool isEqual(const VTableSlot &LHS,
263  const VTableSlot &RHS) {
264  return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset;
265  }
266 };
267 
268 } // end namespace llvm
269 
270 namespace {
271 
272 // A virtual call site. VTable is the loaded virtual table pointer, and CS is
273 // the indirect virtual call.
274 struct VirtualCallSite {
275  Value *VTable;
276  CallSite CS;
277 
278  // If non-null, this field points to the associated unsafe use count stored in
279  // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description
280  // of that field for details.
281  unsigned *NumUnsafeUses;
282 
283  void
284  emitRemark(const StringRef OptName, const StringRef TargetName,
286  Function *F = CS.getCaller();
287  DebugLoc DLoc = CS->getDebugLoc();
288  BasicBlock *Block = CS.getParent();
289 
290  using namespace ore;
291  OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, OptName, DLoc, Block)
292  << NV("Optimization", OptName)
293  << ": devirtualized a call to "
294  << NV("FunctionName", TargetName));
295  }
296 
297  void replaceAndErase(
298  const StringRef OptName, const StringRef TargetName, bool RemarksEnabled,
300  Value *New) {
301  if (RemarksEnabled)
302  emitRemark(OptName, TargetName, OREGetter);
303  CS->replaceAllUsesWith(New);
304  if (auto II = dyn_cast<InvokeInst>(CS.getInstruction())) {
305  BranchInst::Create(II->getNormalDest(), CS.getInstruction());
306  II->getUnwindDest()->removePredecessor(II->getParent());
307  }
308  CS->eraseFromParent();
309  // This use is no longer unsafe.
310  if (NumUnsafeUses)
311  --*NumUnsafeUses;
312  }
313 };
314 
315 // Call site information collected for a specific VTableSlot and possibly a list
316 // of constant integer arguments. The grouping by arguments is handled by the
317 // VTableSlotInfo class.
318 struct CallSiteInfo {
319  /// The set of call sites for this slot. Used during regular LTO and the
320  /// import phase of ThinLTO (as well as the export phase of ThinLTO for any
321  /// call sites that appear in the merged module itself); in each of these
322  /// cases we are directly operating on the call sites at the IR level.
323  std::vector<VirtualCallSite> CallSites;
324 
325  /// Whether all call sites represented by this CallSiteInfo, including those
326  /// in summaries, have been devirtualized. This starts off as true because a
327  /// default constructed CallSiteInfo represents no call sites.
328  bool AllCallSitesDevirted = true;
329 
330  // These fields are used during the export phase of ThinLTO and reflect
331  // information collected from function summaries.
332 
333  /// Whether any function summary contains an llvm.assume(llvm.type.test) for
334  /// this slot.
335  bool SummaryHasTypeTestAssumeUsers = false;
336 
337  /// CFI-specific: a vector containing the list of function summaries that use
338  /// the llvm.type.checked.load intrinsic and therefore will require
339  /// resolutions for llvm.type.test in order to implement CFI checks if
340  /// devirtualization was unsuccessful. If devirtualization was successful, the
341  /// pass will clear this vector by calling markDevirt(). If at the end of the
342  /// pass the vector is non-empty, we will need to add a use of llvm.type.test
343  /// to each of the function summaries in the vector.
344  std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers;
345 
346  bool isExported() const {
347  return SummaryHasTypeTestAssumeUsers ||
348  !SummaryTypeCheckedLoadUsers.empty();
349  }
350 
351  void markSummaryHasTypeTestAssumeUsers() {
352  SummaryHasTypeTestAssumeUsers = true;
353  AllCallSitesDevirted = false;
354  }
355 
356  void addSummaryTypeCheckedLoadUser(FunctionSummary *FS) {
357  SummaryTypeCheckedLoadUsers.push_back(FS);
358  AllCallSitesDevirted = false;
359  }
360 
361  void markDevirt() {
362  AllCallSitesDevirted = true;
363 
364  // As explained in the comment for SummaryTypeCheckedLoadUsers.
365  SummaryTypeCheckedLoadUsers.clear();
366  }
367 };
368 
369 // Call site information collected for a specific VTableSlot.
370 struct VTableSlotInfo {
371  // The set of call sites which do not have all constant integer arguments
372  // (excluding "this").
373  CallSiteInfo CSInfo;
374 
375  // The set of call sites with all constant integer arguments (excluding
376  // "this"), grouped by argument list.
377  std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo;
378 
379  void addCallSite(Value *VTable, CallSite CS, unsigned *NumUnsafeUses);
380 
381 private:
382  CallSiteInfo &findCallSiteInfo(CallSite CS);
383 };
384 
385 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallSite CS) {
386  std::vector<uint64_t> Args;
387  auto *CI = dyn_cast<IntegerType>(CS.getType());
388  if (!CI || CI->getBitWidth() > 64 || CS.arg_empty())
389  return CSInfo;
390  for (auto &&Arg : make_range(CS.arg_begin() + 1, CS.arg_end())) {
391  auto *CI = dyn_cast<ConstantInt>(Arg);
392  if (!CI || CI->getBitWidth() > 64)
393  return CSInfo;
394  Args.push_back(CI->getZExtValue());
395  }
396  return ConstCSInfo[Args];
397 }
398 
399 void VTableSlotInfo::addCallSite(Value *VTable, CallSite CS,
400  unsigned *NumUnsafeUses) {
401  auto &CSI = findCallSiteInfo(CS);
402  CSI.AllCallSitesDevirted = false;
403  CSI.CallSites.push_back({VTable, CS, NumUnsafeUses});
404 }
405 
406 struct DevirtModule {
407  Module &M;
409 
410  ModuleSummaryIndex *ExportSummary;
411  const ModuleSummaryIndex *ImportSummary;
412 
413  IntegerType *Int8Ty;
414  PointerType *Int8PtrTy;
416  IntegerType *Int64Ty;
417  IntegerType *IntPtrTy;
418 
419  bool RemarksEnabled;
421 
423 
424  // This map keeps track of the number of "unsafe" uses of a loaded function
425  // pointer. The key is the associated llvm.type.test intrinsic call generated
426  // by this pass. An unsafe use is one that calls the loaded function pointer
427  // directly. Every time we eliminate an unsafe use (for example, by
428  // devirtualizing it or by applying virtual constant propagation), we
429  // decrement the value stored in this map. If a value reaches zero, we can
430  // eliminate the type check by RAUWing the associated llvm.type.test call with
431  // true.
432  std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest;
433 
434  DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter,
436  ModuleSummaryIndex *ExportSummary,
437  const ModuleSummaryIndex *ImportSummary)
438  : M(M), AARGetter(AARGetter), ExportSummary(ExportSummary),
439  ImportSummary(ImportSummary), Int8Ty(Type::getInt8Ty(M.getContext())),
440  Int8PtrTy(Type::getInt8PtrTy(M.getContext())),
442  Int64Ty(Type::getInt64Ty(M.getContext())),
443  IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)),
444  RemarksEnabled(areRemarksEnabled()), OREGetter(OREGetter) {
445  assert(!(ExportSummary && ImportSummary));
446  }
447 
448  bool areRemarksEnabled();
449 
450  void scanTypeTestUsers(Function *TypeTestFunc, Function *AssumeFunc);
451  void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc);
452 
453  void buildTypeIdentifierMap(
454  std::vector<VTableBits> &Bits,
455  DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap);
456  Constant *getPointerAtOffset(Constant *I, uint64_t Offset);
457  bool
458  tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot,
459  const std::set<TypeMemberInfo> &TypeMemberInfos,
460  uint64_t ByteOffset);
461 
462  void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn,
463  bool &IsExported);
464  bool trySingleImplDevirt(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
465  VTableSlotInfo &SlotInfo,
467 
468  void applyICallBranchFunnel(VTableSlotInfo &SlotInfo, Constant *JT,
469  bool &IsExported);
470  void tryICallBranchFunnel(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
471  VTableSlotInfo &SlotInfo,
472  WholeProgramDevirtResolution *Res, VTableSlot Slot);
473 
474  bool tryEvaluateFunctionsWithArgs(
475  MutableArrayRef<VirtualCallTarget> TargetsForSlot,
477 
478  void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
479  uint64_t TheRetVal);
480  bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
481  CallSiteInfo &CSInfo,
483 
484  // Returns the global symbol name that is used to export information about the
485  // given vtable slot and list of arguments.
486  std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args,
487  StringRef Name);
488 
489  bool shouldExportConstantsAsAbsoluteSymbols();
490 
491  // This function is called during the export phase to create a symbol
492  // definition containing information about the given vtable slot and list of
493  // arguments.
494  void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name,
495  Constant *C);
496  void exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name,
497  uint32_t Const, uint32_t &Storage);
498 
499  // This function is called during the import phase to create a reference to
500  // the symbol definition created during the export phase.
501  Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
502  StringRef Name);
503  Constant *importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
504  StringRef Name, IntegerType *IntTy,
505  uint32_t Storage);
506 
507  Constant *getMemberAddr(const TypeMemberInfo *M);
508 
509  void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne,
510  Constant *UniqueMemberAddr);
511  bool tryUniqueRetValOpt(unsigned BitWidth,
512  MutableArrayRef<VirtualCallTarget> TargetsForSlot,
513  CallSiteInfo &CSInfo,
515  VTableSlot Slot, ArrayRef<uint64_t> Args);
516 
517  void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName,
518  Constant *Byte, Constant *Bit);
519  bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
520  VTableSlotInfo &SlotInfo,
521  WholeProgramDevirtResolution *Res, VTableSlot Slot);
522 
523  void rebuildGlobal(VTableBits &B);
524 
525  // Apply the summary resolution for Slot to all virtual calls in SlotInfo.
526  void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo);
527 
528  // If we were able to eliminate all unsafe uses for a type checked load,
529  // eliminate the associated type tests by replacing them with true.
530  void removeRedundantTypeTests();
531 
532  bool run();
533 
534  // Lower the module using the action and summary passed as command line
535  // arguments. For testing purposes only.
536  static bool runForTesting(
537  Module &M, function_ref<AAResults &(Function &)> AARGetter,
539 };
540 
541 struct WholeProgramDevirt : public ModulePass {
542  static char ID;
543 
544  bool UseCommandLine = false;
545 
546  ModuleSummaryIndex *ExportSummary;
547  const ModuleSummaryIndex *ImportSummary;
548 
549  WholeProgramDevirt() : ModulePass(ID), UseCommandLine(true) {
551  }
552 
553  WholeProgramDevirt(ModuleSummaryIndex *ExportSummary,
554  const ModuleSummaryIndex *ImportSummary)
555  : ModulePass(ID), ExportSummary(ExportSummary),
556  ImportSummary(ImportSummary) {
558  }
559 
560  bool runOnModule(Module &M) override {
561  if (skipModule(M))
562  return false;
563 
564  // In the new pass manager, we can request the optimization
565  // remark emitter pass on a per-function-basis, which the
566  // OREGetter will do for us.
567  // In the old pass manager, this is harder, so we just build
568  // an optimization remark emitter on the fly, when we need it.
569  std::unique_ptr<OptimizationRemarkEmitter> ORE;
570  auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & {
571  ORE = make_unique<OptimizationRemarkEmitter>(F);
572  return *ORE;
573  };
574 
575  if (UseCommandLine)
576  return DevirtModule::runForTesting(M, LegacyAARGetter(*this), OREGetter);
577 
578  return DevirtModule(M, LegacyAARGetter(*this), OREGetter, ExportSummary,
579  ImportSummary)
580  .run();
581  }
582 
583  void getAnalysisUsage(AnalysisUsage &AU) const override {
586  }
587 };
588 
589 } // end anonymous namespace
590 
591 INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt",
592  "Whole program devirtualization", false, false)
596  "Whole program devirtualization", false, false)
597 char WholeProgramDevirt::ID = 0;
598 
599 ModulePass *
601  const ModuleSummaryIndex *ImportSummary) {
602  return new WholeProgramDevirt(ExportSummary, ImportSummary);
603 }
604 
606  ModuleAnalysisManager &AM) {
607  auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
608  auto AARGetter = [&](Function &F) -> AAResults & {
609  return FAM.getResult<AAManager>(F);
610  };
611  auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & {
612  return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F);
613  };
614  if (!DevirtModule(M, AARGetter, OREGetter, nullptr, nullptr).run())
615  return PreservedAnalyses::all();
616  return PreservedAnalyses::none();
617 }
618 
619 bool DevirtModule::runForTesting(
620  Module &M, function_ref<AAResults &(Function &)> AARGetter,
622  ModuleSummaryIndex Summary(/*IsPerformingAnalysis=*/false);
623 
624  // Handle the command-line summary arguments. This code is for testing
625  // purposes only, so we handle errors directly.
626  if (!ClReadSummary.empty()) {
627  ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary +
628  ": ");
629  auto ReadSummaryFile =
631 
632  yaml::Input In(ReadSummaryFile->getBuffer());
633  In >> Summary;
634  ExitOnErr(errorCodeToError(In.error()));
635  }
636 
637  bool Changed =
638  DevirtModule(
639  M, AARGetter, OREGetter,
640  ClSummaryAction == PassSummaryAction::Export ? &Summary : nullptr,
641  ClSummaryAction == PassSummaryAction::Import ? &Summary : nullptr)
642  .run();
643 
644  if (!ClWriteSummary.empty()) {
645  ExitOnError ExitOnErr(
646  "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": ");
647  std::error_code EC;
649  ExitOnErr(errorCodeToError(EC));
650 
651  yaml::Output Out(OS);
652  Out << Summary;
653  }
654 
655  return Changed;
656 }
657 
658 void DevirtModule::buildTypeIdentifierMap(
659  std::vector<VTableBits> &Bits,
660  DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) {
662  Bits.reserve(M.getGlobalList().size());
664  for (GlobalVariable &GV : M.globals()) {
665  Types.clear();
666  GV.getMetadata(LLVMContext::MD_type, Types);
667  if (Types.empty())
668  continue;
669 
670  VTableBits *&BitsPtr = GVToBits[&GV];
671  if (!BitsPtr) {
672  Bits.emplace_back();
673  Bits.back().GV = &GV;
674  Bits.back().ObjectSize =
675  M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType());
676  BitsPtr = &Bits.back();
677  }
678 
679  for (MDNode *Type : Types) {
680  auto TypeID = Type->getOperand(1).get();
681 
682  uint64_t Offset =
683  cast<ConstantInt>(
684  cast<ConstantAsMetadata>(Type->getOperand(0))->getValue())
685  ->getZExtValue();
686 
687  TypeIdMap[TypeID].insert({BitsPtr, Offset});
688  }
689  }
690 }
691 
692 Constant *DevirtModule::getPointerAtOffset(Constant *I, uint64_t Offset) {
693  if (I->getType()->isPointerTy()) {
694  if (Offset == 0)
695  return I;
696  return nullptr;
697  }
698 
699  const DataLayout &DL = M.getDataLayout();
700 
701  if (auto *C = dyn_cast<ConstantStruct>(I)) {
702  const StructLayout *SL = DL.getStructLayout(C->getType());
703  if (Offset >= SL->getSizeInBytes())
704  return nullptr;
705 
706  unsigned Op = SL->getElementContainingOffset(Offset);
707  return getPointerAtOffset(cast<Constant>(I->getOperand(Op)),
708  Offset - SL->getElementOffset(Op));
709  }
710  if (auto *C = dyn_cast<ConstantArray>(I)) {
711  ArrayType *VTableTy = C->getType();
712  uint64_t ElemSize = DL.getTypeAllocSize(VTableTy->getElementType());
713 
714  unsigned Op = Offset / ElemSize;
715  if (Op >= C->getNumOperands())
716  return nullptr;
717 
718  return getPointerAtOffset(cast<Constant>(I->getOperand(Op)),
719  Offset % ElemSize);
720  }
721  return nullptr;
722 }
723 
724 bool DevirtModule::tryFindVirtualCallTargets(
725  std::vector<VirtualCallTarget> &TargetsForSlot,
726  const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset) {
727  for (const TypeMemberInfo &TM : TypeMemberInfos) {
728  if (!TM.Bits->GV->isConstant())
729  return false;
730 
731  Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(),
732  TM.Offset + ByteOffset);
733  if (!Ptr)
734  return false;
735 
736  auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts());
737  if (!Fn)
738  return false;
739 
740  // We can disregard __cxa_pure_virtual as a possible call target, as
741  // calls to pure virtuals are UB.
742  if (Fn->getName() == "__cxa_pure_virtual")
743  continue;
744 
745  TargetsForSlot.push_back({Fn, &TM});
746  }
747 
748  // Give up if we couldn't find any targets.
749  return !TargetsForSlot.empty();
750 }
751 
752 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo,
753  Constant *TheFn, bool &IsExported) {
754  auto Apply = [&](CallSiteInfo &CSInfo) {
755  for (auto &&VCallSite : CSInfo.CallSites) {
756  if (RemarksEnabled)
757  VCallSite.emitRemark("single-impl", TheFn->getName(), OREGetter);
758  VCallSite.CS.setCalledFunction(ConstantExpr::getBitCast(
759  TheFn, VCallSite.CS.getCalledValue()->getType()));
760  // This use is no longer unsafe.
761  if (VCallSite.NumUnsafeUses)
762  --*VCallSite.NumUnsafeUses;
763  }
764  if (CSInfo.isExported())
765  IsExported = true;
766  CSInfo.markDevirt();
767  };
768  Apply(SlotInfo.CSInfo);
769  for (auto &P : SlotInfo.ConstCSInfo)
770  Apply(P.second);
771 }
772 
773 bool DevirtModule::trySingleImplDevirt(
774  MutableArrayRef<VirtualCallTarget> TargetsForSlot,
775  VTableSlotInfo &SlotInfo, WholeProgramDevirtResolution *Res) {
776  // See if the program contains a single implementation of this virtual
777  // function.
778  Function *TheFn = TargetsForSlot[0].Fn;
779  for (auto &&Target : TargetsForSlot)
780  if (TheFn != Target.Fn)
781  return false;
782 
783  // If so, update each call site to call that implementation directly.
784  if (RemarksEnabled)
785  TargetsForSlot[0].WasDevirt = true;
786 
787  bool IsExported = false;
788  applySingleImplDevirt(SlotInfo, TheFn, IsExported);
789  if (!IsExported)
790  return false;
791 
792  // If the only implementation has local linkage, we must promote to external
793  // to make it visible to thin LTO objects. We can only get here during the
794  // ThinLTO export phase.
795  if (TheFn->hasLocalLinkage()) {
796  std::string NewName = (TheFn->getName() + "$merged").str();
797 
798  // Since we are renaming the function, any comdats with the same name must
799  // also be renamed. This is required when targeting COFF, as the comdat name
800  // must match one of the names of the symbols in the comdat.
801  if (Comdat *C = TheFn->getComdat()) {
802  if (C->getName() == TheFn->getName()) {
803  Comdat *NewC = M.getOrInsertComdat(NewName);
804  NewC->setSelectionKind(C->getSelectionKind());
805  for (GlobalObject &GO : M.global_objects())
806  if (GO.getComdat() == C)
807  GO.setComdat(NewC);
808  }
809  }
810 
813  TheFn->setName(NewName);
814  }
815 
817  Res->SingleImplName = TheFn->getName();
818 
819  return true;
820 }
821 
822 void DevirtModule::tryICallBranchFunnel(
823  MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo,
824  WholeProgramDevirtResolution *Res, VTableSlot Slot) {
825  Triple T(M.getTargetTriple());
826  if (T.getArch() != Triple::x86_64)
827  return;
828 
829  if (TargetsForSlot.size() > ClThreshold)
830  return;
831 
832  bool HasNonDevirt = !SlotInfo.CSInfo.AllCallSitesDevirted;
833  if (!HasNonDevirt)
834  for (auto &P : SlotInfo.ConstCSInfo)
835  if (!P.second.AllCallSitesDevirted) {
836  HasNonDevirt = true;
837  break;
838  }
839 
840  if (!HasNonDevirt)
841  return;
842 
843  FunctionType *FT =
844  FunctionType::get(Type::getVoidTy(M.getContext()), {Int8PtrTy}, true);
845  Function *JT;
846  if (isa<MDString>(Slot.TypeID)) {
848  getGlobalName(Slot, {}, "branch_funnel"), &M);
849  JT->setVisibility(GlobalValue::HiddenVisibility);
850  } else {
851  JT = Function::Create(FT, Function::InternalLinkage, "branch_funnel", &M);
852  }
853  JT->addAttribute(1, Attribute::Nest);
854 
855  std::vector<Value *> JTArgs;
856  JTArgs.push_back(JT->arg_begin());
857  for (auto &T : TargetsForSlot) {
858  JTArgs.push_back(getMemberAddr(T.TM));
859  JTArgs.push_back(T.Fn);
860  }
861 
862  BasicBlock *BB = BasicBlock::Create(M.getContext(), "", JT, nullptr);
863  Constant *Intr =
864  Intrinsic::getDeclaration(&M, llvm::Intrinsic::icall_branch_funnel, {});
865 
866  auto *CI = CallInst::Create(Intr, JTArgs, "", BB);
867  CI->setTailCallKind(CallInst::TCK_MustTail);
868  ReturnInst::Create(M.getContext(), nullptr, BB);
869 
870  bool IsExported = false;
871  applyICallBranchFunnel(SlotInfo, JT, IsExported);
872  if (IsExported)
874 }
875 
876 void DevirtModule::applyICallBranchFunnel(VTableSlotInfo &SlotInfo,
877  Constant *JT, bool &IsExported) {
878  auto Apply = [&](CallSiteInfo &CSInfo) {
879  if (CSInfo.isExported())
880  IsExported = true;
881  if (CSInfo.AllCallSitesDevirted)
882  return;
883  for (auto &&VCallSite : CSInfo.CallSites) {
884  CallSite CS = VCallSite.CS;
885 
886  // Jump tables are only profitable if the retpoline mitigation is enabled.
887  Attribute FSAttr = CS.getCaller()->getFnAttribute("target-features");
888  if (FSAttr.hasAttribute(Attribute::None) ||
889  !FSAttr.getValueAsString().contains("+retpoline"))
890  continue;
891 
892  if (RemarksEnabled)
893  VCallSite.emitRemark("branch-funnel", JT->getName(), OREGetter);
894 
895  // Pass the address of the vtable in the nest register, which is r10 on
896  // x86_64.
897  std::vector<Type *> NewArgs;
898  NewArgs.push_back(Int8PtrTy);
899  for (Type *T : CS.getFunctionType()->params())
900  NewArgs.push_back(T);
903  CS.getFunctionType()->isVarArg()));
904 
905  IRBuilder<> IRB(CS.getInstruction());
906  std::vector<Value *> Args;
907  Args.push_back(IRB.CreateBitCast(VCallSite.VTable, Int8PtrTy));
908  for (unsigned I = 0; I != CS.getNumArgOperands(); ++I)
909  Args.push_back(CS.getArgOperand(I));
910 
911  CallSite NewCS;
912  if (CS.isCall())
913  NewCS = IRB.CreateCall(IRB.CreateBitCast(JT, NewFT), Args);
914  else
915  NewCS = IRB.CreateInvoke(
916  IRB.CreateBitCast(JT, NewFT),
917  cast<InvokeInst>(CS.getInstruction())->getNormalDest(),
918  cast<InvokeInst>(CS.getInstruction())->getUnwindDest(), Args);
919  NewCS.setCallingConv(CS.getCallingConv());
920 
922  std::vector<AttributeSet> NewArgAttrs;
923  NewArgAttrs.push_back(AttributeSet::get(
925  M.getContext(), Attribute::Nest)}));
926  for (unsigned I = 0; I + 2 < Attrs.getNumAttrSets(); ++I)
927  NewArgAttrs.push_back(Attrs.getParamAttributes(I));
928  NewCS.setAttributes(
930  Attrs.getRetAttributes(), NewArgAttrs));
931 
932  CS->replaceAllUsesWith(NewCS.getInstruction());
933  CS->eraseFromParent();
934 
935  // This use is no longer unsafe.
936  if (VCallSite.NumUnsafeUses)
937  --*VCallSite.NumUnsafeUses;
938  }
939  // Don't mark as devirtualized because there may be callers compiled without
940  // retpoline mitigation, which would mean that they are lowered to
941  // llvm.type.test and therefore require an llvm.type.test resolution for the
942  // type identifier.
943  };
944  Apply(SlotInfo.CSInfo);
945  for (auto &P : SlotInfo.ConstCSInfo)
946  Apply(P.second);
947 }
948 
949 bool DevirtModule::tryEvaluateFunctionsWithArgs(
950  MutableArrayRef<VirtualCallTarget> TargetsForSlot,
952  // Evaluate each function and store the result in each target's RetVal
953  // field.
954  for (VirtualCallTarget &Target : TargetsForSlot) {
955  if (Target.Fn->arg_size() != Args.size() + 1)
956  return false;
957 
958  Evaluator Eval(M.getDataLayout(), nullptr);
960  EvalArgs.push_back(
961  Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0)));
962  for (unsigned I = 0; I != Args.size(); ++I) {
963  auto *ArgTy = dyn_cast<IntegerType>(
964  Target.Fn->getFunctionType()->getParamType(I + 1));
965  if (!ArgTy)
966  return false;
967  EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I]));
968  }
969 
970  Constant *RetVal;
971  if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) ||
972  !isa<ConstantInt>(RetVal))
973  return false;
974  Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue();
975  }
976  return true;
977 }
978 
979 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
980  uint64_t TheRetVal) {
981  for (auto Call : CSInfo.CallSites)
982  Call.replaceAndErase(
983  "uniform-ret-val", FnName, RemarksEnabled, OREGetter,
984  ConstantInt::get(cast<IntegerType>(Call.CS.getType()), TheRetVal));
985  CSInfo.markDevirt();
986 }
987 
988 bool DevirtModule::tryUniformRetValOpt(
989  MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo,
991  // Uniform return value optimization. If all functions return the same
992  // constant, replace all calls with that constant.
993  uint64_t TheRetVal = TargetsForSlot[0].RetVal;
994  for (const VirtualCallTarget &Target : TargetsForSlot)
995  if (Target.RetVal != TheRetVal)
996  return false;
997 
998  if (CSInfo.isExported()) {
1000  Res->Info = TheRetVal;
1001  }
1002 
1003  applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal);
1004  if (RemarksEnabled)
1005  for (auto &&Target : TargetsForSlot)
1006  Target.WasDevirt = true;
1007  return true;
1008 }
1009 
1010 std::string DevirtModule::getGlobalName(VTableSlot Slot,
1011  ArrayRef<uint64_t> Args,
1012  StringRef Name) {
1013  std::string FullName = "__typeid_";
1014  raw_string_ostream OS(FullName);
1015  OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset;
1016  for (uint64_t Arg : Args)
1017  OS << '_' << Arg;
1018  OS << '_' << Name;
1019  return OS.str();
1020 }
1021 
1022 bool DevirtModule::shouldExportConstantsAsAbsoluteSymbols() {
1023  Triple T(M.getTargetTriple());
1024  return (T.getArch() == Triple::x86 || T.getArch() == Triple::x86_64) &&
1025  T.getObjectFormat() == Triple::ELF;
1026 }
1027 
1028 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
1029  StringRef Name, Constant *C) {
1031  getGlobalName(Slot, Args, Name), C, &M);
1033 }
1034 
1035 void DevirtModule::exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
1036  StringRef Name, uint32_t Const,
1037  uint32_t &Storage) {
1038  if (shouldExportConstantsAsAbsoluteSymbols()) {
1039  exportGlobal(
1040  Slot, Args, Name,
1041  ConstantExpr::getIntToPtr(ConstantInt::get(Int32Ty, Const), Int8PtrTy));
1042  return;
1043  }
1044 
1045  Storage = Const;
1046 }
1047 
1048 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
1049  StringRef Name) {
1050  Constant *C = M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Ty);
1051  auto *GV = dyn_cast<GlobalVariable>(C);
1052  if (GV)
1054  return C;
1055 }
1056 
1057 Constant *DevirtModule::importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
1058  StringRef Name, IntegerType *IntTy,
1059  uint32_t Storage) {
1060  if (!shouldExportConstantsAsAbsoluteSymbols())
1061  return ConstantInt::get(IntTy, Storage);
1062 
1063  Constant *C = importGlobal(Slot, Args, Name);
1064  auto *GV = cast<GlobalVariable>(C->stripPointerCasts());
1065  C = ConstantExpr::getPtrToInt(C, IntTy);
1066 
1067  // We only need to set metadata if the global is newly created, in which
1068  // case it would not have hidden visibility.
1069  if (GV->getMetadata(LLVMContext::MD_absolute_symbol))
1070  return C;
1071 
1072  auto SetAbsRange = [&](uint64_t Min, uint64_t Max) {
1073  auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min));
1074  auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max));
1075  GV->setMetadata(LLVMContext::MD_absolute_symbol,
1076  MDNode::get(M.getContext(), {MinC, MaxC}));
1077  };
1078  unsigned AbsWidth = IntTy->getBitWidth();
1079  if (AbsWidth == IntPtrTy->getBitWidth())
1080  SetAbsRange(~0ull, ~0ull); // Full set.
1081  else
1082  SetAbsRange(0, 1ull << AbsWidth);
1083  return C;
1084 }
1085 
1086 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
1087  bool IsOne,
1088  Constant *UniqueMemberAddr) {
1089  for (auto &&Call : CSInfo.CallSites) {
1090  IRBuilder<> B(Call.CS.getInstruction());
1091  Value *Cmp =
1092  B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
1093  B.CreateBitCast(Call.VTable, Int8PtrTy), UniqueMemberAddr);
1094  Cmp = B.CreateZExt(Cmp, Call.CS->getType());
1095  Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter,
1096  Cmp);
1097  }
1098  CSInfo.markDevirt();
1099 }
1100 
1101 Constant *DevirtModule::getMemberAddr(const TypeMemberInfo *M) {
1102  Constant *C = ConstantExpr::getBitCast(M->Bits->GV, Int8PtrTy);
1103  return ConstantExpr::getGetElementPtr(Int8Ty, C,
1104  ConstantInt::get(Int64Ty, M->Offset));
1105 }
1106 
1107 bool DevirtModule::tryUniqueRetValOpt(
1108  unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot,
1109  CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res,
1110  VTableSlot Slot, ArrayRef<uint64_t> Args) {
1111  // IsOne controls whether we look for a 0 or a 1.
1112  auto tryUniqueRetValOptFor = [&](bool IsOne) {
1113  const TypeMemberInfo *UniqueMember = nullptr;
1114  for (const VirtualCallTarget &Target : TargetsForSlot) {
1115  if (Target.RetVal == (IsOne ? 1 : 0)) {
1116  if (UniqueMember)
1117  return false;
1118  UniqueMember = Target.TM;
1119  }
1120  }
1121 
1122  // We should have found a unique member or bailed out by now. We already
1123  // checked for a uniform return value in tryUniformRetValOpt.
1124  assert(UniqueMember);
1125 
1126  Constant *UniqueMemberAddr = getMemberAddr(UniqueMember);
1127  if (CSInfo.isExported()) {
1129  Res->Info = IsOne;
1130 
1131  exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr);
1132  }
1133 
1134  // Replace each call with the comparison.
1135  applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne,
1136  UniqueMemberAddr);
1137 
1138  // Update devirtualization statistics for targets.
1139  if (RemarksEnabled)
1140  for (auto &&Target : TargetsForSlot)
1141  Target.WasDevirt = true;
1142 
1143  return true;
1144  };
1145 
1146  if (BitWidth == 1) {
1147  if (tryUniqueRetValOptFor(true))
1148  return true;
1149  if (tryUniqueRetValOptFor(false))
1150  return true;
1151  }
1152  return false;
1153 }
1154 
1155 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName,
1156  Constant *Byte, Constant *Bit) {
1157  for (auto Call : CSInfo.CallSites) {
1158  auto *RetType = cast<IntegerType>(Call.CS.getType());
1159  IRBuilder<> B(Call.CS.getInstruction());
1160  Value *Addr =
1161  B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte);
1162  if (RetType->getBitWidth() == 1) {
1163  Value *Bits = B.CreateLoad(Addr);
1164  Value *BitsAndBit = B.CreateAnd(Bits, Bit);
1165  auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0));
1166  Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled,
1167  OREGetter, IsBitSet);
1168  } else {
1169  Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo());
1170  Value *Val = B.CreateLoad(RetType, ValAddr);
1171  Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled,
1172  OREGetter, Val);
1173  }
1174  }
1175  CSInfo.markDevirt();
1176 }
1177 
1178 bool DevirtModule::tryVirtualConstProp(
1179  MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo,
1180  WholeProgramDevirtResolution *Res, VTableSlot Slot) {
1181  // This only works if the function returns an integer.
1182  auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType());
1183  if (!RetType)
1184  return false;
1185  unsigned BitWidth = RetType->getBitWidth();
1186  if (BitWidth > 64)
1187  return false;
1188 
1189  // Make sure that each function is defined, does not access memory, takes at
1190  // least one argument, does not use its first argument (which we assume is
1191  // 'this'), and has the same return type.
1192  //
1193  // Note that we test whether this copy of the function is readnone, rather
1194  // than testing function attributes, which must hold for any copy of the
1195  // function, even a less optimized version substituted at link time. This is
1196  // sound because the virtual constant propagation optimizations effectively
1197  // inline all implementations of the virtual function into each call site,
1198  // rather than using function attributes to perform local optimization.
1199  for (VirtualCallTarget &Target : TargetsForSlot) {
1200  if (Target.Fn->isDeclaration() ||
1201  computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) !=
1202  MAK_ReadNone ||
1203  Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() ||
1204  Target.Fn->getReturnType() != RetType)
1205  return false;
1206  }
1207 
1208  for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) {
1209  if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first))
1210  continue;
1211 
1212  WholeProgramDevirtResolution::ByArg *ResByArg = nullptr;
1213  if (Res)
1214  ResByArg = &Res->ResByArg[CSByConstantArg.first];
1215 
1216  if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg))
1217  continue;
1218 
1219  if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second,
1220  ResByArg, Slot, CSByConstantArg.first))
1221  continue;
1222 
1223  // Find an allocation offset in bits in all vtables associated with the
1224  // type.
1225  uint64_t AllocBefore =
1226  findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth);
1227  uint64_t AllocAfter =
1228  findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth);
1229 
1230  // Calculate the total amount of padding needed to store a value at both
1231  // ends of the object.
1232  uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0;
1233  for (auto &&Target : TargetsForSlot) {
1234  TotalPaddingBefore += std::max<int64_t>(
1235  (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0);
1236  TotalPaddingAfter += std::max<int64_t>(
1237  (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0);
1238  }
1239 
1240  // If the amount of padding is too large, give up.
1241  // FIXME: do something smarter here.
1242  if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128)
1243  continue;
1244 
1245  // Calculate the offset to the value as a (possibly negative) byte offset
1246  // and (if applicable) a bit offset, and store the values in the targets.
1247  int64_t OffsetByte;
1248  uint64_t OffsetBit;
1249  if (TotalPaddingBefore <= TotalPaddingAfter)
1250  setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte,
1251  OffsetBit);
1252  else
1253  setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte,
1254  OffsetBit);
1255 
1256  if (RemarksEnabled)
1257  for (auto &&Target : TargetsForSlot)
1258  Target.WasDevirt = true;
1259 
1260 
1261  if (CSByConstantArg.second.isExported()) {
1263  exportConstant(Slot, CSByConstantArg.first, "byte", OffsetByte,
1264  ResByArg->Byte);
1265  exportConstant(Slot, CSByConstantArg.first, "bit", 1ULL << OffsetBit,
1266  ResByArg->Bit);
1267  }
1268 
1269  // Rewrite each call to a load from OffsetByte/OffsetBit.
1270  Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte);
1271  Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit);
1272  applyVirtualConstProp(CSByConstantArg.second,
1273  TargetsForSlot[0].Fn->getName(), ByteConst, BitConst);
1274  }
1275  return true;
1276 }
1277 
1278 void DevirtModule::rebuildGlobal(VTableBits &B) {
1279  if (B.Before.Bytes.empty() && B.After.Bytes.empty())
1280  return;
1281 
1282  // Align each byte array to pointer width.
1283  unsigned PointerSize = M.getDataLayout().getPointerSize();
1284  B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), PointerSize));
1285  B.After.Bytes.resize(alignTo(B.After.Bytes.size(), PointerSize));
1286 
1287  // Before was stored in reverse order; flip it now.
1288  for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I)
1289  std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]);
1290 
1291  // Build an anonymous global containing the before bytes, followed by the
1292  // original initializer, followed by the after bytes.
1293  auto NewInit = ConstantStruct::getAnon(
1294  {ConstantDataArray::get(M.getContext(), B.Before.Bytes),
1295  B.GV->getInitializer(),
1296  ConstantDataArray::get(M.getContext(), B.After.Bytes)});
1297  auto NewGV =
1298  new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(),
1299  GlobalVariable::PrivateLinkage, NewInit, "", B.GV);
1300  NewGV->setSection(B.GV->getSection());
1301  NewGV->setComdat(B.GV->getComdat());
1302 
1303  // Copy the original vtable's metadata to the anonymous global, adjusting
1304  // offsets as required.
1305  NewGV->copyMetadata(B.GV, B.Before.Bytes.size());
1306 
1307  // Build an alias named after the original global, pointing at the second
1308  // element (the original initializer).
1309  auto Alias = GlobalAlias::create(
1310  B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "",
1312  NewInit->getType(), NewGV,
1314  ConstantInt::get(Int32Ty, 1)}),
1315  &M);
1316  Alias->setVisibility(B.GV->getVisibility());
1317  Alias->takeName(B.GV);
1318 
1319  B.GV->replaceAllUsesWith(Alias);
1320  B.GV->eraseFromParent();
1321 }
1322 
1323 bool DevirtModule::areRemarksEnabled() {
1324  const auto &FL = M.getFunctionList();
1325  if (FL.empty())
1326  return false;
1327  const Function &Fn = FL.front();
1328 
1329  const auto &BBL = Fn.getBasicBlockList();
1330  if (BBL.empty())
1331  return false;
1332  auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front());
1333  return DI.isEnabled();
1334 }
1335 
1336 void DevirtModule::scanTypeTestUsers(Function *TypeTestFunc,
1337  Function *AssumeFunc) {
1338  // Find all virtual calls via a virtual table pointer %p under an assumption
1339  // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p
1340  // points to a member of the type identifier %md. Group calls by (type ID,
1341  // offset) pair (effectively the identity of the virtual function) and store
1342  // to CallSlots.
1343  DenseSet<Value *> SeenPtrs;
1344  for (auto I = TypeTestFunc->use_begin(), E = TypeTestFunc->use_end();
1345  I != E;) {
1346  auto CI = dyn_cast<CallInst>(I->getUser());
1347  ++I;
1348  if (!CI)
1349  continue;
1350 
1351  // Search for virtual calls based on %p and add them to DevirtCalls.
1352  SmallVector<DevirtCallSite, 1> DevirtCalls;
1354  findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI);
1355 
1356  // If we found any, add them to CallSlots. Only do this if we haven't seen
1357  // the vtable pointer before, as it may have been CSE'd with pointers from
1358  // other call sites, and we don't want to process call sites multiple times.
1359  if (!Assumes.empty()) {
1360  Metadata *TypeId =
1361  cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata();
1362  Value *Ptr = CI->getArgOperand(0)->stripPointerCasts();
1363  if (SeenPtrs.insert(Ptr).second) {
1364  for (DevirtCallSite Call : DevirtCalls) {
1365  CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS, nullptr);
1366  }
1367  }
1368  }
1369 
1370  // We no longer need the assumes or the type test.
1371  for (auto Assume : Assumes)
1372  Assume->eraseFromParent();
1373  // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we
1374  // may use the vtable argument later.
1375  if (CI->use_empty())
1376  CI->eraseFromParent();
1377  }
1378 }
1379 
1380 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) {
1381  Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test);
1382 
1383  for (auto I = TypeCheckedLoadFunc->use_begin(),
1384  E = TypeCheckedLoadFunc->use_end();
1385  I != E;) {
1386  auto CI = dyn_cast<CallInst>(I->getUser());
1387  ++I;
1388  if (!CI)
1389  continue;
1390 
1391  Value *Ptr = CI->getArgOperand(0);
1392  Value *Offset = CI->getArgOperand(1);
1393  Value *TypeIdValue = CI->getArgOperand(2);
1394  Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata();
1395 
1396  SmallVector<DevirtCallSite, 1> DevirtCalls;
1397  SmallVector<Instruction *, 1> LoadedPtrs;
1399  bool HasNonCallUses = false;
1400  findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds,
1401  HasNonCallUses, CI);
1402 
1403  // Start by generating "pessimistic" code that explicitly loads the function
1404  // pointer from the vtable and performs the type check. If possible, we will
1405  // eliminate the load and the type check later.
1406 
1407  // If possible, only generate the load at the point where it is used.
1408  // This helps avoid unnecessary spills.
1409  IRBuilder<> LoadB(
1410  (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI);
1411  Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset);
1412  Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy));
1413  Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr);
1414 
1415  for (Instruction *LoadedPtr : LoadedPtrs) {
1416  LoadedPtr->replaceAllUsesWith(LoadedValue);
1417  LoadedPtr->eraseFromParent();
1418  }
1419 
1420  // Likewise for the type test.
1421  IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI);
1422  CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue});
1423 
1424  for (Instruction *Pred : Preds) {
1425  Pred->replaceAllUsesWith(TypeTestCall);
1426  Pred->eraseFromParent();
1427  }
1428 
1429  // We have already erased any extractvalue instructions that refer to the
1430  // intrinsic call, but the intrinsic may have other non-extractvalue uses
1431  // (although this is unlikely). In that case, explicitly build a pair and
1432  // RAUW it.
1433  if (!CI->use_empty()) {
1434  Value *Pair = UndefValue::get(CI->getType());
1435  IRBuilder<> B(CI);
1436  Pair = B.CreateInsertValue(Pair, LoadedValue, {0});
1437  Pair = B.CreateInsertValue(Pair, TypeTestCall, {1});
1438  CI->replaceAllUsesWith(Pair);
1439  }
1440 
1441  // The number of unsafe uses is initially the number of uses.
1442  auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall];
1443  NumUnsafeUses = DevirtCalls.size();
1444 
1445  // If the function pointer has a non-call user, we cannot eliminate the type
1446  // check, as one of those users may eventually call the pointer. Increment
1447  // the unsafe use count to make sure it cannot reach zero.
1448  if (HasNonCallUses)
1449  ++NumUnsafeUses;
1450  for (DevirtCallSite Call : DevirtCalls) {
1451  CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS,
1452  &NumUnsafeUses);
1453  }
1454 
1455  CI->eraseFromParent();
1456  }
1457 }
1458 
1459 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) {
1460  const TypeIdSummary *TidSummary =
1461  ImportSummary->getTypeIdSummary(cast<MDString>(Slot.TypeID)->getString());
1462  if (!TidSummary)
1463  return;
1464  auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset);
1465  if (ResI == TidSummary->WPDRes.end())
1466  return;
1467  const WholeProgramDevirtResolution &Res = ResI->second;
1468 
1470  // The type of the function in the declaration is irrelevant because every
1471  // call site will cast it to the correct type.
1472  auto *SingleImpl = M.getOrInsertFunction(
1473  Res.SingleImplName, Type::getVoidTy(M.getContext()));
1474 
1475  // This is the import phase so we should not be exporting anything.
1476  bool IsExported = false;
1477  applySingleImplDevirt(SlotInfo, SingleImpl, IsExported);
1478  assert(!IsExported);
1479  }
1480 
1481  for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) {
1482  auto I = Res.ResByArg.find(CSByConstantArg.first);
1483  if (I == Res.ResByArg.end())
1484  continue;
1485  auto &ResByArg = I->second;
1486  // FIXME: We should figure out what to do about the "function name" argument
1487  // to the apply* functions, as the function names are unavailable during the
1488  // importing phase. For now we just pass the empty string. This does not
1489  // impact correctness because the function names are just used for remarks.
1490  switch (ResByArg.TheKind) {
1492  applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info);
1493  break;
1495  Constant *UniqueMemberAddr =
1496  importGlobal(Slot, CSByConstantArg.first, "unique_member");
1497  applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info,
1498  UniqueMemberAddr);
1499  break;
1500  }
1502  Constant *Byte = importConstant(Slot, CSByConstantArg.first, "byte",
1503  Int32Ty, ResByArg.Byte);
1504  Constant *Bit = importConstant(Slot, CSByConstantArg.first, "bit", Int8Ty,
1505  ResByArg.Bit);
1506  applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit);
1507  break;
1508  }
1509  default:
1510  break;
1511  }
1512  }
1513 
1515  auto *JT = M.getOrInsertFunction(getGlobalName(Slot, {}, "branch_funnel"),
1516  Type::getVoidTy(M.getContext()));
1517  bool IsExported = false;
1518  applyICallBranchFunnel(SlotInfo, JT, IsExported);
1519  assert(!IsExported);
1520  }
1521 }
1522 
1523 void DevirtModule::removeRedundantTypeTests() {
1524  auto True = ConstantInt::getTrue(M.getContext());
1525  for (auto &&U : NumUnsafeUsesForTypeTest) {
1526  if (U.second == 0) {
1527  U.first->replaceAllUsesWith(True);
1528  U.first->eraseFromParent();
1529  }
1530  }
1531 }
1532 
1533 bool DevirtModule::run() {
1534  Function *TypeTestFunc =
1535  M.getFunction(Intrinsic::getName(Intrinsic::type_test));
1536  Function *TypeCheckedLoadFunc =
1537  M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load));
1538  Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume));
1539 
1540  // Normally if there are no users of the devirtualization intrinsics in the
1541  // module, this pass has nothing to do. But if we are exporting, we also need
1542  // to handle any users that appear only in the function summaries.
1543  if (!ExportSummary &&
1544  (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc ||
1545  AssumeFunc->use_empty()) &&
1546  (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty()))
1547  return false;
1548 
1549  if (TypeTestFunc && AssumeFunc)
1550  scanTypeTestUsers(TypeTestFunc, AssumeFunc);
1551 
1552  if (TypeCheckedLoadFunc)
1553  scanTypeCheckedLoadUsers(TypeCheckedLoadFunc);
1554 
1555  if (ImportSummary) {
1556  for (auto &S : CallSlots)
1557  importResolution(S.first, S.second);
1558 
1559  removeRedundantTypeTests();
1560 
1561  // The rest of the code is only necessary when exporting or during regular
1562  // LTO, so we are done.
1563  return true;
1564  }
1565 
1566  // Rebuild type metadata into a map for easy lookup.
1567  std::vector<VTableBits> Bits;
1569  buildTypeIdentifierMap(Bits, TypeIdMap);
1570  if (TypeIdMap.empty())
1571  return true;
1572 
1573  // Collect information from summary about which calls to try to devirtualize.
1574  if (ExportSummary) {
1576  for (auto &P : TypeIdMap) {
1577  if (auto *TypeId = dyn_cast<MDString>(P.first))
1578  MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back(
1579  TypeId);
1580  }
1581 
1582  for (auto &P : *ExportSummary) {
1583  for (auto &S : P.second.SummaryList) {
1584  auto *FS = dyn_cast<FunctionSummary>(S.get());
1585  if (!FS)
1586  continue;
1587  // FIXME: Only add live functions.
1588  for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) {
1589  for (Metadata *MD : MetadataByGUID[VF.GUID]) {
1590  CallSlots[{MD, VF.Offset}]
1591  .CSInfo.markSummaryHasTypeTestAssumeUsers();
1592  }
1593  }
1594  for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) {
1595  for (Metadata *MD : MetadataByGUID[VF.GUID]) {
1596  CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS);
1597  }
1598  }
1599  for (const FunctionSummary::ConstVCall &VC :
1600  FS->type_test_assume_const_vcalls()) {
1601  for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) {
1602  CallSlots[{MD, VC.VFunc.Offset}]
1603  .ConstCSInfo[VC.Args]
1604  .markSummaryHasTypeTestAssumeUsers();
1605  }
1606  }
1607  for (const FunctionSummary::ConstVCall &VC :
1608  FS->type_checked_load_const_vcalls()) {
1609  for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) {
1610  CallSlots[{MD, VC.VFunc.Offset}]
1611  .ConstCSInfo[VC.Args]
1612  .addSummaryTypeCheckedLoadUser(FS);
1613  }
1614  }
1615  }
1616  }
1617  }
1618 
1619  // For each (type, offset) pair:
1620  bool DidVirtualConstProp = false;
1621  std::map<std::string, Function*> DevirtTargets;
1622  for (auto &S : CallSlots) {
1623  // Search each of the members of the type identifier for the virtual
1624  // function implementation at offset S.first.ByteOffset, and add to
1625  // TargetsForSlot.
1626  std::vector<VirtualCallTarget> TargetsForSlot;
1627  if (tryFindVirtualCallTargets(TargetsForSlot, TypeIdMap[S.first.TypeID],
1628  S.first.ByteOffset)) {
1629  WholeProgramDevirtResolution *Res = nullptr;
1630  if (ExportSummary && isa<MDString>(S.first.TypeID))
1631  Res = &ExportSummary
1632  ->getOrInsertTypeIdSummary(
1633  cast<MDString>(S.first.TypeID)->getString())
1634  .WPDRes[S.first.ByteOffset];
1635 
1636  if (!trySingleImplDevirt(TargetsForSlot, S.second, Res)) {
1637  DidVirtualConstProp |=
1638  tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first);
1639 
1640  tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first);
1641  }
1642 
1643  // Collect functions devirtualized at least for one call site for stats.
1644  if (RemarksEnabled)
1645  for (const auto &T : TargetsForSlot)
1646  if (T.WasDevirt)
1647  DevirtTargets[T.Fn->getName()] = T.Fn;
1648  }
1649 
1650  // CFI-specific: if we are exporting and any llvm.type.checked.load
1651  // intrinsics were *not* devirtualized, we need to add the resulting
1652  // llvm.type.test intrinsics to the function summaries so that the
1653  // LowerTypeTests pass will export them.
1654  if (ExportSummary && isa<MDString>(S.first.TypeID)) {
1655  auto GUID =
1656  GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString());
1657  for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers)
1658  FS->addTypeTest(GUID);
1659  for (auto &CCS : S.second.ConstCSInfo)
1660  for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers)
1661  FS->addTypeTest(GUID);
1662  }
1663  }
1664 
1665  if (RemarksEnabled) {
1666  // Generate remarks for each devirtualized function.
1667  for (const auto &DT : DevirtTargets) {
1668  Function *F = DT.second;
1669 
1670  using namespace ore;
1671  OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F)
1672  << "devirtualized "
1673  << NV("FunctionName", F->getName()));
1674  }
1675  }
1676 
1677  removeRedundantTypeTests();
1678 
1679  // Rebuild each global we touched as part of virtual constant propagation to
1680  // include the before and after bytes.
1681  if (DidVirtualConstProp)
1682  for (VTableBits &B : Bits)
1683  rebuildGlobal(B);
1684 
1685  return true;
1686 }
void setVisibility(VisibilityTypes V)
Definition: GlobalValue.h:238
uint64_t CallInst * C
StringRef getSection() const
Get the custom section of this global if it has one.
Definition: GlobalObject.h:90
void push_back(const T &Elt)
Definition: SmallVector.h:212
use_iterator use_end()
Definition: Value.h:352
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:111
const std::string & getTargetTriple() const
Get the target triple which is a string describing the target host.
Definition: Module.h:233
bool empty() const
Definition: Function.h:643
bool hasLocalLinkage() const
Definition: GlobalValue.h:435
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
DiagnosticInfoOptimizationBase::Argument NV
CallingConv::ID getCallingConv() const
Get the calling convention of the call.
Definition: CallSite.h:312
Whole program devirtualization
static unsigned getHashValue(const VTableSlot &I)
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:687
const Constant * getInitializer() const
getInitializer - Return the initializer for this global variable.
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE bool contains(StringRef Other) const
Return true if the given string is a substring of *this, and false otherwise.
Definition: StringRef.h:448
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:63
static Constant * getGetElementPtr(Type *Ty, Constant *C, ArrayRef< Constant *> IdxList, bool InBounds=false, Optional< unsigned > InRangeIndex=None, Type *OnlyIfReducedTy=nullptr)
Getelementptr form.
Definition: Constants.h:1128
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM)
const StructLayout * getStructLayout(StructType *Ty) const
Returns a StructLayout object, indicating the alignment of the struct, its size, and the offsets of i...
Definition: DataLayout.cpp:588
This provides a very simple, boring adaptor for a begin and end iterator into a range type...
Helper for check-and-exit error handling.
Definition: Error.h:1120
void initializeWholeProgramDevirtPass(PassRegistry &)
This class represents a function call, abstracting a target machine&#39;s calling convention.
This file contains the declarations for metadata subclasses.
An immutable pass that tracks lazily created AssumptionCache objects.
An efficient, type-erasing, non-owning reference to a callable.
Definition: STLExtras.h:93
static Constant * getIntToPtr(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1727
uint64_t Info
Additional information for the resolution:
Like Internal, but omit from symbol table.
Definition: GlobalValue.h:57
Externally visible function.
Definition: GlobalValue.h:49
static cl::opt< unsigned > ClThreshold("wholeprogramdevirt-branch-funnel-threshold", cl::Hidden, cl::init(10), cl::ZeroOrMore, cl::desc("Maximum number of call targets per " "call site to enable branch funnels"))
This class implements a map that also provides access to all stored values in a deterministic order...
Definition: MapVector.h:38
A debug info location.
Definition: DebugLoc.h:34
Metadata node.
Definition: Metadata.h:862
F(f)
static CallInst * Create(Value *Func, ArrayRef< Value *> Args, ArrayRef< OperandBundleDef > Bundles=None, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
uint64_t alignTo(uint64_t Value, uint64_t Align, uint64_t Skew=0)
Returns the next integer (mod 2**64) that is greater than or equal to Value and is a multiple of Alig...
Definition: MathExtras.h:677
const GlobalListType & getGlobalList() const
Get the Module&#39;s list of global variables (constant).
Definition: Module.h:498
static IntegerType * getInt64Ty(LLVMContext &C)
Definition: Type.cpp:177
Hexagon Common GEP
MemoryAccessKind computeFunctionBodyMemoryAccess(Function &F, AAResults &AAR)
Returns the memory access properties of this copy of the function.
void setAfterReturnValues(MutableArrayRef< VirtualCallTarget > Targets, uint64_t AllocAfter, unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit)
std::map< uint64_t, WholeProgramDevirtResolution > WPDRes
Mapping from byte offset to whole-program devirt resolution for that (typeid, byte offset) pair...
unsigned getElementContainingOffset(uint64_t Offset) const
Given a valid byte offset into the structure, returns the structure index that contains it...
Definition: DataLayout.cpp:84
static Constant * getNullValue(Type *Ty)
Constructor to create a &#39;0&#39; constant of arbitrary type.
Definition: Constants.cpp:258
Export information to summary.
const char * getName() const
getName - Get the target name.
iterator_range< global_object_iterator > global_objects()
Definition: Module.h:636
bool arg_empty() const
Definition: CallSite.h:218
static ReturnInst * Create(LLVMContext &C, Value *retVal=nullptr, Instruction *InsertBefore=nullptr)
void setBeforeReturnValues(MutableArrayRef< VirtualCallTarget > Targets, uint64_t AllocBefore, unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit)
void findDevirtualizableCallsForTypeTest(SmallVectorImpl< DevirtCallSite > &DevirtCalls, SmallVectorImpl< CallInst *> &Assumes, const CallInst *CI)
Given a call to the intrinsic .type.test, find all devirtualizable call sites based on the call and r...
AnalysisUsage & addRequired()
Used to lazily calculate structure layout information for a target machine, based on the DataLayout s...
Definition: DataLayout.h:521
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
StringRef getName(ID id)
Return the LLVM name for an intrinsic, such as "llvm.ppc.altivec.lvx".
Definition: Function.cpp:604
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:361
When retpoline mitigation is enabled, use a branch funnel that is defined in the merged module...
LLVMContext & getContext() const
Get the global data context.
Definition: Module.h:237
The returned value is undefined.
Definition: MathExtras.h:38
bool isConstant() const
If the value is a global constant, its value is immutable throughout the runtime execution of the pro...
IterTy arg_end() const
Definition: CallSite.h:575
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:677
This class is a functor to be used in legacy module or SCC passes for computing AA results for a func...
TypeID
Definitions of all of the base types for the Type system.
Definition: Type.h:55
No attributes have been set.
Definition: Attributes.h:71
AttributeSet getRetAttributes() const
The attributes for the ret value are returned.
InstrTy * getInstruction() const
Definition: CallSite.h:92
unsigned Intr
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:295
static bool isEqual(const VTableSlot &LHS, const VTableSlot &RHS)
enum llvm::WholeProgramDevirtResolution::Kind TheKind
Class to represent function types.
Definition: DerivedTypes.h:103
Value * CreateBitCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1570
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
#define T
ValTy * getArgOperand(unsigned i) const
Definition: CallSite.h:297
Class to represent array types.
Definition: DerivedTypes.h:369
AttributeSet getParamAttributes(unsigned ArgNo) const
The attributes for the argument or parameter at the given index are returned.
bool isVarArg() const
Definition: DerivedTypes.h:123
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:33
LinkageTypes getLinkage() const
Definition: GlobalValue.h:450
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:439
constexpr char Attrs[]
Key for Kernel::Metadata::mAttrs.
unsigned getBitWidth() const
Get the number of bits in this IntegerType.
Definition: DerivedTypes.h:66
Class to hold module path string table and global value map, and encapsulate methods for operating on...
static ConstantAsMetadata * get(Constant *C)
Definition: Metadata.h:408
Function * getDeclaration(Module *M, ID id, ArrayRef< Type *> Tys=None)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:993
Value * getOperand(unsigned i) const
Definition: User.h:170
Class to represent pointers.
Definition: DerivedTypes.h:467
static PreservedAnalyses none()
Convenience factory function for the empty preserved set.
Definition: PassManager.h:156
bool isCall() const
Return true if a CallInst is enclosed.
Definition: CallSite.h:87
static Constant * getBitCast(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1740
LoadInst * CreateLoad(Value *Ptr, const char *Name)
Provided to resolve &#39;CreateLoad(Ptr, "...")&#39; correctly, instead of converting the string to &#39;bool&#39; fo...
Definition: IRBuilder.h:1251
IntegerType * getIntPtrType(LLVMContext &C, unsigned AddressSpace=0) const
Returns an integer type with size at least as big as that of a pointer in the given address space...
Definition: DataLayout.cpp:744
static MDTuple * get(LLVMContext &Context, ArrayRef< Metadata *> MDs)
Definition: Metadata.h:1164
void setAttributes(AttributeList PAL)
Set the parameter attributes of the call.
Definition: CallSite.h:333
#define P(N)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:406
unsigned getNumOperands() const
bool hasAttribute(AttrKind Val) const
Return true if the attribute is present.
Definition: Attributes.cpp:201
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
static AttributeSet get(LLVMContext &C, const AttrBuilder &B)
Definition: Attributes.cpp:510
std::size_t countTrailingZeros(T Val, ZeroBehavior ZB=ZB_Width)
Count number of 0&#39;s from the least significant bit to the most stopping at the first 1...
Definition: MathExtras.h:112
VisibilityTypes getVisibility() const
Definition: GlobalValue.h:232
MutableArrayRef - Represent a mutable reference to an array (0 or more elements consecutively in memo...
Definition: ArrayRef.h:291
Import information from summary.
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
A call site that could be devirtualized.
Constant * getOrInsertGlobal(StringRef Name, Type *Ty)
Look up the specified global in the module symbol table.
Definition: Module.cpp:202
size_t size() const
size - Get the array size.
Definition: ArrayRef.h:149
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
Error errorCodeToError(std::error_code EC)
Helper for converting an std::error_code to a Error.
Definition: Error.cpp:78
ValuesClass values(OptsTy... Options)
Helper to build a ValuesClass by forwarding a variable number of arguments as an initializer list to ...
Definition: CommandLine.h:626
PreservedAnalyses run(Module &M, ModuleAnalysisManager &)
This file contains the declarations for the subclasses of Constant, which represent the different fla...
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:221
#define DEBUG_TYPE
A manager for alias analyses.
std::pair< iterator, bool > insert(const ValueT &V)
Definition: DenseSet.h:187
Diagnostic information for applied optimization remarks.
void eraseFromParent()
eraseFromParent - This method unlinks &#39;this&#39; from the containing module and deletes it...
Definition: Globals.cpp:354
INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt", "Whole program devirtualization", false, false) INITIALIZE_PASS_END(WholeProgramDevirt
ArrayRef< Type * > params() const
Definition: DerivedTypes.h:130
Expected< T > errorOrToExpected(ErrorOr< T > &&EO)
Convert an ErrorOr<T> to an Expected<T>.
Definition: Error.h:1082
Represent the analysis usage information of a pass.
static Type * getVoidTy(LLVMContext &C)
Definition: Type.cpp:161
Type::TypeID TypeID
void setCallingConv(CallingConv::ID CC)
Set the calling convention of the call.
Definition: CallSite.h:316
static FunctionType * get(Type *Result, ArrayRef< Type *> Params, bool isVarArg)
This static method is the primary way of constructing a FunctionType.
Definition: Type.cpp:297
static BasicBlock * Create(LLVMContext &Context, const Twine &Name="", Function *Parent=nullptr, BasicBlock *InsertBefore=nullptr)
Creates a new BasicBlock.
Definition: BasicBlock.h:101
void reserve(size_type NumEntries)
Grow the densemap so that it can contain at least NumEntries items before resizing again...
Definition: DenseMap.h:100
Class to represent integer types.
Definition: DerivedTypes.h:40
enum llvm::WholeProgramDevirtResolution::ByArg::Kind TheKind
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1382
const Constant * stripPointerCasts() const
Definition: Constant.h:169
Comdat * getOrInsertComdat(StringRef Name)
Return the Comdat in the module with the specified name.
Definition: Module.cpp:467
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:567
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
unsigned getNumArgOperands() const
Definition: CallSite.h:293
static PointerType * getInt8PtrTy(LLVMContext &C, unsigned AS=0)
Definition: Type.cpp:220
bool isEnabled() const override
std::string & str()
Flushes the stream contents to the target string and returns the string&#39;s reference.
Definition: raw_ostream.h:483
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
Triple - Helper class for working with autoconf configuration names.
Definition: Triple.h:44
An "identifier" for a virtual function.
This class evaluates LLVM IR, producing the Constant representing each SSA instruction.
Definition: Evaluator.h:38
Value * CreateGEP(Value *Ptr, ArrayRef< Value *> IdxList, const Twine &Name="")
Definition: IRBuilder.h:1320
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
static PointerType * getUnqual(Type *ElementType)
This constructs a pointer to an object of the specified type in the generic address space (address sp...
Definition: DerivedTypes.h:482
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
void setSelectionKind(SelectionKind Val)
Definition: Comdat.h:46
void findDevirtualizableCallsForTypeCheckedLoad(SmallVectorImpl< DevirtCallSite > &DevirtCalls, SmallVectorImpl< Instruction *> &LoadedPtrs, SmallVectorImpl< Instruction *> &Preds, bool &HasNonCallUses, const CallInst *CI)
Given a call to the intrinsic .type.checked.load, find all devirtualizable call sites based on the ca...
IterTy arg_begin() const
Definition: CallSite.h:571
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:862
Module.h This file contains the declarations for the Module class.
Single implementation devirtualization.
Type * getReturnType() const
Definition: DerivedTypes.h:124
uint64_t getSizeInBytes() const
Definition: DataLayout.h:529
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:611
static BranchInst * Create(BasicBlock *IfTrue, Instruction *InsertBefore=nullptr)
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:567
BBTy * getParent() const
Get the basic block containing the call site.
Definition: CallSite.h:97
void setLinkage(LinkageTypes LT)
Definition: GlobalValue.h:444
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:924
wholeprogramdevirt
Target - Wrapper for Target specific information.
unsigned getNumAttrSets() const
A specification for a virtual function call with all constant integer arguments.
GUID getGUID() const
Return a 64-bit global unique ID constructed from global value name (i.e.
Definition: GlobalValue.h:500
The file should be opened in text mode on platforms that make this distinction.
Definition: FileSystem.h:688
ArrayRef< T > slice(size_t N, size_t M) const
slice(n, m) - Chop off the first N elements of the array, and keep M elements in the array...
Definition: ArrayRef.h:179
#define clEnumValN(ENUMVAL, FLAGNAME, DESC)
Definition: CommandLine.h:601
amdgpu Simplify well known AMD library false Value Value * Arg
const Comdat * getComdat() const
Definition: GlobalObject.h:101
ModulePass * createWholeProgramDevirtPass(ModuleSummaryIndex *ExportSummary, const ModuleSummaryIndex *ImportSummary)
This pass implements whole-program devirtualization using type metadata.
uint64_t getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:428
use_iterator use_begin()
Definition: Value.h:344
FunTy * getCaller() const
Return the caller function for this call site.
Definition: CallSite.h:267
std::map< std::vector< uint64_t >, ByArg > ResByArg
Resolutions for calls with all constant integer arguments (excluding the first argument, "this"), where the key is the argument vector.
A raw_ostream that writes to a file descriptor.
Definition: raw_ostream.h:364
uint64_t getElementOffset(unsigned Idx) const
Definition: DataLayout.h:543
static IntegerType * getInt32Ty(LLVMContext &C)
Definition: Type.cpp:176
uint64_t findLowestOffset(ArrayRef< VirtualCallTarget > Targets, bool IsAfter, uint64_t Size)
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:61
StringRef getValueAsString() const
Return the attribute&#39;s value as a string.
Definition: Attributes.cpp:194
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:224
static Constant * getPtrToInt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1714
#define I(x, y, z)
Definition: MD5.cpp:58
ModulePass class - This class is used to implement unstructured interprocedural optimizations and ana...
Definition: Pass.h:225
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
const BasicBlockListType & getBasicBlockList() const
Get the underlying elements of the Function...
Definition: Function.h:614
Rename collisions when linking (static functions).
Definition: GlobalValue.h:56
static Attribute get(LLVMContext &Context, AttrKind Kind, uint64_t Val=0)
Return a uniquified Attribute object.
Definition: Attributes.cpp:80
Provides passes for computing function attributes based on interprocedural analyses.
Function summary information to aid decisions and implementation of importing.
Type * getType() const
Return the type of the instruction that generated this call site.
Definition: CallSite.h:264
static ErrorOr< std::unique_ptr< MemoryBuffer > > getFile(const Twine &Filename, int64_t FileSize=-1, bool RequiresNullTerminator=true, bool IsVolatile=false)
Open the specified file as a MemoryBuffer, returning a new MemoryBuffer if successful, otherwise returning null.
static cl::opt< PassSummaryAction > ClSummaryAction("wholeprogramdevirt-summary-action", cl::desc("What to do with the summary when running this pass"), cl::values(clEnumValN(PassSummaryAction::None, "none", "Do nothing"), clEnumValN(PassSummaryAction::Import, "import", "Import typeid resolutions from summary and globals"), clEnumValN(PassSummaryAction::Export, "export", "Export typeid resolutions to summary and globals")), cl::Hidden)
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
const BasicBlock & front() const
Definition: Function.h:644
A raw_ostream that writes to an std::string.
Definition: raw_ostream.h:467
aarch64 promote const
LLVM Value Representation.
Definition: Value.h:73
FunctionType * getFunctionType() const
Definition: CallSite.h:320
constexpr char Size[]
Key for Kernel::Arg::Metadata::mSize.
static const Function * getParent(const Value *V)
AttributeSet getFnAttributes() const
The function attributes are returned.
Attribute getFnAttribute(Attribute::AttrKind Kind) const
Return the attribute for the given attribute kind.
Definition: Function.h:312
static Constant * getAnon(ArrayRef< Constant *> V, bool Packed=false)
Return an anonymous struct that has the specified elements.
Definition: Constants.h:465
static Constant * get(LLVMContext &Context, ArrayRef< ElementTy > Elts)
get() constructor - Return a constant with array type with an element count and element type matching...
Definition: Constants.h:699
Type * getElementType() const
Definition: DerivedTypes.h:360
iterator_range< global_iterator > globals()
Definition: Module.h:561
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
This is the interface for LLVM&#39;s primary stateless and local alias analysis.
static cl::opt< std::string > ClReadSummary("wholeprogramdevirt-read-summary", cl::desc("Read summary from given YAML file before running pass"), cl::Hidden)
static cl::opt< std::string > ClWriteSummary("wholeprogramdevirt-write-summary", cl::desc("Write summary to given YAML file after running pass"), cl::Hidden)
A container for analyses that lazily runs them and caches their results.
static Function * Create(FunctionType *Ty, LinkageTypes Linkage, const Twine &N="", Module *M=nullptr)
Definition: Function.h:136
AttributeList getAttributes() const
Get the parameter attributes of the call.
Definition: CallSite.h:329
Root of the metadata hierarchy.
Definition: Metadata.h:58
static IntegerType * getInt8Ty(LLVMContext &C)
Definition: Type.cpp:174
void setSection(StringRef S)
Change the section for this global.
Definition: Globals.cpp:184
The optimization diagnostic interface.
bool use_empty() const
Definition: Value.h:328
static GlobalAlias * create(Type *Ty, unsigned AddressSpace, LinkageTypes Linkage, const Twine &Name, Constant *Aliasee, Module *Parent)
If a parent module is specified, the alias is automatically inserted into the end of the specified mo...
Definition: Globals.cpp:418
constexpr char Args[]
Key for Kernel::Metadata::mArgs.
static AttributeList get(LLVMContext &C, ArrayRef< std::pair< unsigned, Attribute >> Attrs)
Create an AttributeList with the specified parameters in it.
Definition: Attributes.cpp:870
IntegerType * Int32Ty
CallInst * CreateCall(Value *Callee, ArrayRef< Value *> Args=None, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1812
An analysis over an "outer" IR unit that provides access to an analysis manager over an "inner" IR un...
Definition: PassManager.h:946