LLVM  8.0.0svn
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/Dominators.h"
62 #include "llvm/IR/Function.h"
63 #include "llvm/IR/GlobalAlias.h"
64 #include "llvm/IR/GlobalVariable.h"
65 #include "llvm/IR/IRBuilder.h"
66 #include "llvm/IR/InstrTypes.h"
67 #include "llvm/IR/Instruction.h"
68 #include "llvm/IR/Instructions.h"
69 #include "llvm/IR/Intrinsics.h"
70 #include "llvm/IR/LLVMContext.h"
71 #include "llvm/IR/Metadata.h"
72 #include "llvm/IR/Module.h"
74 #include "llvm/Pass.h"
75 #include "llvm/PassRegistry.h"
76 #include "llvm/PassSupport.h"
77 #include "llvm/Support/Casting.h"
78 #include "llvm/Support/Error.h"
81 #include "llvm/Transforms/IPO.h"
84 #include <algorithm>
85 #include <cstddef>
86 #include <map>
87 #include <set>
88 #include <string>
89 
90 using namespace llvm;
91 using namespace wholeprogramdevirt;
92 
93 #define DEBUG_TYPE "wholeprogramdevirt"
94 
96  "wholeprogramdevirt-summary-action",
97  cl::desc("What to do with the summary when running this pass"),
98  cl::values(clEnumValN(PassSummaryAction::None, "none", "Do nothing"),
100  "Import typeid resolutions from summary and globals"),
102  "Export typeid resolutions to summary and globals")),
103  cl::Hidden);
104 
106  "wholeprogramdevirt-read-summary",
107  cl::desc("Read summary from given YAML file before running pass"),
108  cl::Hidden);
109 
111  "wholeprogramdevirt-write-summary",
112  cl::desc("Write summary to given YAML file after running pass"),
113  cl::Hidden);
114 
115 static cl::opt<unsigned>
116  ClThreshold("wholeprogramdevirt-branch-funnel-threshold", cl::Hidden,
118  cl::desc("Maximum number of call targets per "
119  "call site to enable branch funnels"));
120 
121 // Find the minimum offset that we may store a value of size Size bits at. If
122 // IsAfter is set, look for an offset before the object, otherwise look for an
123 // offset after the object.
124 uint64_t
126  bool IsAfter, uint64_t Size) {
127  // Find a minimum offset taking into account only vtable sizes.
128  uint64_t MinByte = 0;
129  for (const VirtualCallTarget &Target : Targets) {
130  if (IsAfter)
131  MinByte = std::max(MinByte, Target.minAfterBytes());
132  else
133  MinByte = std::max(MinByte, Target.minBeforeBytes());
134  }
135 
136  // Build a vector of arrays of bytes covering, for each target, a slice of the
137  // used region (see AccumBitVector::BytesUsed in
138  // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively,
139  // this aligns the used regions to start at MinByte.
140  //
141  // In this example, A, B and C are vtables, # is a byte already allocated for
142  // a virtual function pointer, AAAA... (etc.) are the used regions for the
143  // vtables and Offset(X) is the value computed for the Offset variable below
144  // for X.
145  //
146  // Offset(A)
147  // | |
148  // |MinByte
149  // A: ################AAAAAAAA|AAAAAAAA
150  // B: ########BBBBBBBBBBBBBBBB|BBBB
151  // C: ########################|CCCCCCCCCCCCCCCC
152  // | Offset(B) |
153  //
154  // This code produces the slices of A, B and C that appear after the divider
155  // at MinByte.
156  std::vector<ArrayRef<uint8_t>> Used;
157  for (const VirtualCallTarget &Target : Targets) {
158  ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed
159  : Target.TM->Bits->Before.BytesUsed;
160  uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes()
161  : MinByte - Target.minBeforeBytes();
162 
163  // Disregard used regions that are smaller than Offset. These are
164  // effectively all-free regions that do not need to be checked.
165  if (VTUsed.size() > Offset)
166  Used.push_back(VTUsed.slice(Offset));
167  }
168 
169  if (Size == 1) {
170  // Find a free bit in each member of Used.
171  for (unsigned I = 0;; ++I) {
172  uint8_t BitsUsed = 0;
173  for (auto &&B : Used)
174  if (I < B.size())
175  BitsUsed |= B[I];
176  if (BitsUsed != 0xff)
177  return (MinByte + I) * 8 +
178  countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined);
179  }
180  } else {
181  // Find a free (Size/8) byte region in each member of Used.
182  // FIXME: see if alignment helps.
183  for (unsigned I = 0;; ++I) {
184  for (auto &&B : Used) {
185  unsigned Byte = 0;
186  while ((I + Byte) < B.size() && Byte < (Size / 8)) {
187  if (B[I + Byte])
188  goto NextI;
189  ++Byte;
190  }
191  }
192  return (MinByte + I) * 8;
193  NextI:;
194  }
195  }
196 }
197 
199  MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore,
200  unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
201  if (BitWidth == 1)
202  OffsetByte = -(AllocBefore / 8 + 1);
203  else
204  OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8);
205  OffsetBit = AllocBefore % 8;
206 
207  for (VirtualCallTarget &Target : Targets) {
208  if (BitWidth == 1)
209  Target.setBeforeBit(AllocBefore);
210  else
211  Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8);
212  }
213 }
214 
216  MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter,
217  unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
218  if (BitWidth == 1)
219  OffsetByte = AllocAfter / 8;
220  else
221  OffsetByte = (AllocAfter + 7) / 8;
222  OffsetBit = AllocAfter % 8;
223 
224  for (VirtualCallTarget &Target : Targets) {
225  if (BitWidth == 1)
226  Target.setAfterBit(AllocAfter);
227  else
228  Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8);
229  }
230 }
231 
233  : Fn(Fn), TM(TM),
234  IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {}
235 
236 namespace {
237 
238 // A slot in a set of virtual tables. The TypeID identifies the set of virtual
239 // tables, and the ByteOffset is the offset in bytes from the address point to
240 // the virtual function pointer.
241 struct VTableSlot {
242  Metadata *TypeID;
243  uint64_t ByteOffset;
244 };
245 
246 } // end anonymous namespace
247 
248 namespace llvm {
249 
250 template <> struct DenseMapInfo<VTableSlot> {
251  static VTableSlot getEmptyKey() {
254  }
255  static VTableSlot getTombstoneKey() {
258  }
259  static unsigned getHashValue(const VTableSlot &I) {
260  return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^
262  }
263  static bool isEqual(const VTableSlot &LHS,
264  const VTableSlot &RHS) {
265  return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset;
266  }
267 };
268 
269 } // end namespace llvm
270 
271 namespace {
272 
273 // A virtual call site. VTable is the loaded virtual table pointer, and CS is
274 // the indirect virtual call.
275 struct VirtualCallSite {
276  Value *VTable;
277  CallSite CS;
278 
279  // If non-null, this field points to the associated unsafe use count stored in
280  // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description
281  // of that field for details.
282  unsigned *NumUnsafeUses;
283 
284  void
285  emitRemark(const StringRef OptName, const StringRef TargetName,
287  Function *F = CS.getCaller();
288  DebugLoc DLoc = CS->getDebugLoc();
289  BasicBlock *Block = CS.getParent();
290 
291  using namespace ore;
292  OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, OptName, DLoc, Block)
293  << NV("Optimization", OptName)
294  << ": devirtualized a call to "
295  << NV("FunctionName", TargetName));
296  }
297 
298  void replaceAndErase(
299  const StringRef OptName, const StringRef TargetName, bool RemarksEnabled,
301  Value *New) {
302  if (RemarksEnabled)
303  emitRemark(OptName, TargetName, OREGetter);
304  CS->replaceAllUsesWith(New);
305  if (auto II = dyn_cast<InvokeInst>(CS.getInstruction())) {
306  BranchInst::Create(II->getNormalDest(), CS.getInstruction());
307  II->getUnwindDest()->removePredecessor(II->getParent());
308  }
309  CS->eraseFromParent();
310  // This use is no longer unsafe.
311  if (NumUnsafeUses)
312  --*NumUnsafeUses;
313  }
314 };
315 
316 // Call site information collected for a specific VTableSlot and possibly a list
317 // of constant integer arguments. The grouping by arguments is handled by the
318 // VTableSlotInfo class.
319 struct CallSiteInfo {
320  /// The set of call sites for this slot. Used during regular LTO and the
321  /// import phase of ThinLTO (as well as the export phase of ThinLTO for any
322  /// call sites that appear in the merged module itself); in each of these
323  /// cases we are directly operating on the call sites at the IR level.
324  std::vector<VirtualCallSite> CallSites;
325 
326  /// Whether all call sites represented by this CallSiteInfo, including those
327  /// in summaries, have been devirtualized. This starts off as true because a
328  /// default constructed CallSiteInfo represents no call sites.
329  bool AllCallSitesDevirted = true;
330 
331  // These fields are used during the export phase of ThinLTO and reflect
332  // information collected from function summaries.
333 
334  /// Whether any function summary contains an llvm.assume(llvm.type.test) for
335  /// this slot.
336  bool SummaryHasTypeTestAssumeUsers = false;
337 
338  /// CFI-specific: a vector containing the list of function summaries that use
339  /// the llvm.type.checked.load intrinsic and therefore will require
340  /// resolutions for llvm.type.test in order to implement CFI checks if
341  /// devirtualization was unsuccessful. If devirtualization was successful, the
342  /// pass will clear this vector by calling markDevirt(). If at the end of the
343  /// pass the vector is non-empty, we will need to add a use of llvm.type.test
344  /// to each of the function summaries in the vector.
345  std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers;
346 
347  bool isExported() const {
348  return SummaryHasTypeTestAssumeUsers ||
349  !SummaryTypeCheckedLoadUsers.empty();
350  }
351 
352  void markSummaryHasTypeTestAssumeUsers() {
353  SummaryHasTypeTestAssumeUsers = true;
354  AllCallSitesDevirted = false;
355  }
356 
357  void addSummaryTypeCheckedLoadUser(FunctionSummary *FS) {
358  SummaryTypeCheckedLoadUsers.push_back(FS);
359  AllCallSitesDevirted = false;
360  }
361 
362  void markDevirt() {
363  AllCallSitesDevirted = true;
364 
365  // As explained in the comment for SummaryTypeCheckedLoadUsers.
366  SummaryTypeCheckedLoadUsers.clear();
367  }
368 };
369 
370 // Call site information collected for a specific VTableSlot.
371 struct VTableSlotInfo {
372  // The set of call sites which do not have all constant integer arguments
373  // (excluding "this").
374  CallSiteInfo CSInfo;
375 
376  // The set of call sites with all constant integer arguments (excluding
377  // "this"), grouped by argument list.
378  std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo;
379 
380  void addCallSite(Value *VTable, CallSite CS, unsigned *NumUnsafeUses);
381 
382 private:
383  CallSiteInfo &findCallSiteInfo(CallSite CS);
384 };
385 
386 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallSite CS) {
387  std::vector<uint64_t> Args;
388  auto *CI = dyn_cast<IntegerType>(CS.getType());
389  if (!CI || CI->getBitWidth() > 64 || CS.arg_empty())
390  return CSInfo;
391  for (auto &&Arg : make_range(CS.arg_begin() + 1, CS.arg_end())) {
392  auto *CI = dyn_cast<ConstantInt>(Arg);
393  if (!CI || CI->getBitWidth() > 64)
394  return CSInfo;
395  Args.push_back(CI->getZExtValue());
396  }
397  return ConstCSInfo[Args];
398 }
399 
400 void VTableSlotInfo::addCallSite(Value *VTable, CallSite CS,
401  unsigned *NumUnsafeUses) {
402  auto &CSI = findCallSiteInfo(CS);
403  CSI.AllCallSitesDevirted = false;
404  CSI.CallSites.push_back({VTable, CS, NumUnsafeUses});
405 }
406 
407 struct DevirtModule {
408  Module &M;
411 
412  ModuleSummaryIndex *ExportSummary;
413  const ModuleSummaryIndex *ImportSummary;
414 
415  IntegerType *Int8Ty;
416  PointerType *Int8PtrTy;
418  IntegerType *Int64Ty;
419  IntegerType *IntPtrTy;
420 
421  bool RemarksEnabled;
423 
425 
426  // This map keeps track of the number of "unsafe" uses of a loaded function
427  // pointer. The key is the associated llvm.type.test intrinsic call generated
428  // by this pass. An unsafe use is one that calls the loaded function pointer
429  // directly. Every time we eliminate an unsafe use (for example, by
430  // devirtualizing it or by applying virtual constant propagation), we
431  // decrement the value stored in this map. If a value reaches zero, we can
432  // eliminate the type check by RAUWing the associated llvm.type.test call with
433  // true.
434  std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest;
435 
436  DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter,
438  function_ref<DominatorTree &(Function &)> LookupDomTree,
439  ModuleSummaryIndex *ExportSummary,
440  const ModuleSummaryIndex *ImportSummary)
441  : M(M), AARGetter(AARGetter), LookupDomTree(LookupDomTree),
442  ExportSummary(ExportSummary), ImportSummary(ImportSummary),
443  Int8Ty(Type::getInt8Ty(M.getContext())),
444  Int8PtrTy(Type::getInt8PtrTy(M.getContext())),
446  Int64Ty(Type::getInt64Ty(M.getContext())),
447  IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)),
448  RemarksEnabled(areRemarksEnabled()), OREGetter(OREGetter) {
449  assert(!(ExportSummary && ImportSummary));
450  }
451 
452  bool areRemarksEnabled();
453 
454  void scanTypeTestUsers(Function *TypeTestFunc, Function *AssumeFunc);
455  void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc);
456 
457  void buildTypeIdentifierMap(
458  std::vector<VTableBits> &Bits,
459  DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap);
460  Constant *getPointerAtOffset(Constant *I, uint64_t Offset);
461  bool
462  tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot,
463  const std::set<TypeMemberInfo> &TypeMemberInfos,
464  uint64_t ByteOffset);
465 
466  void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn,
467  bool &IsExported);
468  bool trySingleImplDevirt(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
469  VTableSlotInfo &SlotInfo,
471 
472  void applyICallBranchFunnel(VTableSlotInfo &SlotInfo, Constant *JT,
473  bool &IsExported);
474  void tryICallBranchFunnel(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
475  VTableSlotInfo &SlotInfo,
476  WholeProgramDevirtResolution *Res, VTableSlot Slot);
477 
478  bool tryEvaluateFunctionsWithArgs(
479  MutableArrayRef<VirtualCallTarget> TargetsForSlot,
481 
482  void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
483  uint64_t TheRetVal);
484  bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
485  CallSiteInfo &CSInfo,
487 
488  // Returns the global symbol name that is used to export information about the
489  // given vtable slot and list of arguments.
490  std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args,
491  StringRef Name);
492 
493  bool shouldExportConstantsAsAbsoluteSymbols();
494 
495  // This function is called during the export phase to create a symbol
496  // definition containing information about the given vtable slot and list of
497  // arguments.
498  void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name,
499  Constant *C);
500  void exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name,
501  uint32_t Const, uint32_t &Storage);
502 
503  // This function is called during the import phase to create a reference to
504  // the symbol definition created during the export phase.
505  Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
506  StringRef Name);
507  Constant *importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
508  StringRef Name, IntegerType *IntTy,
509  uint32_t Storage);
510 
511  Constant *getMemberAddr(const TypeMemberInfo *M);
512 
513  void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne,
514  Constant *UniqueMemberAddr);
515  bool tryUniqueRetValOpt(unsigned BitWidth,
516  MutableArrayRef<VirtualCallTarget> TargetsForSlot,
517  CallSiteInfo &CSInfo,
519  VTableSlot Slot, ArrayRef<uint64_t> Args);
520 
521  void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName,
522  Constant *Byte, Constant *Bit);
523  bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
524  VTableSlotInfo &SlotInfo,
525  WholeProgramDevirtResolution *Res, VTableSlot Slot);
526 
527  void rebuildGlobal(VTableBits &B);
528 
529  // Apply the summary resolution for Slot to all virtual calls in SlotInfo.
530  void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo);
531 
532  // If we were able to eliminate all unsafe uses for a type checked load,
533  // eliminate the associated type tests by replacing them with true.
534  void removeRedundantTypeTests();
535 
536  bool run();
537 
538  // Lower the module using the action and summary passed as command line
539  // arguments. For testing purposes only.
540  static bool
541  runForTesting(Module &M, function_ref<AAResults &(Function &)> AARGetter,
543  function_ref<DominatorTree &(Function &)> LookupDomTree);
544 };
545 
546 struct WholeProgramDevirt : public ModulePass {
547  static char ID;
548 
549  bool UseCommandLine = false;
550 
551  ModuleSummaryIndex *ExportSummary;
552  const ModuleSummaryIndex *ImportSummary;
553 
554  WholeProgramDevirt() : ModulePass(ID), UseCommandLine(true) {
556  }
557 
558  WholeProgramDevirt(ModuleSummaryIndex *ExportSummary,
559  const ModuleSummaryIndex *ImportSummary)
560  : ModulePass(ID), ExportSummary(ExportSummary),
561  ImportSummary(ImportSummary) {
563  }
564 
565  bool runOnModule(Module &M) override {
566  if (skipModule(M))
567  return false;
568 
569  // In the new pass manager, we can request the optimization
570  // remark emitter pass on a per-function-basis, which the
571  // OREGetter will do for us.
572  // In the old pass manager, this is harder, so we just build
573  // an optimization remark emitter on the fly, when we need it.
574  std::unique_ptr<OptimizationRemarkEmitter> ORE;
575  auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & {
576  ORE = make_unique<OptimizationRemarkEmitter>(F);
577  return *ORE;
578  };
579 
580  auto LookupDomTree = [this](Function &F) -> DominatorTree & {
581  return this->getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
582  };
583 
584  if (UseCommandLine)
585  return DevirtModule::runForTesting(M, LegacyAARGetter(*this), OREGetter,
586  LookupDomTree);
587 
588  return DevirtModule(M, LegacyAARGetter(*this), OREGetter, LookupDomTree,
589  ExportSummary, ImportSummary)
590  .run();
591  }
592 
593  void getAnalysisUsage(AnalysisUsage &AU) const override {
597  }
598 };
599 
600 } // end anonymous namespace
601 
602 INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt",
603  "Whole program devirtualization", false, false)
608  "Whole program devirtualization", false, false)
609 char WholeProgramDevirt::ID = 0;
610 
611 ModulePass *
613  const ModuleSummaryIndex *ImportSummary) {
614  return new WholeProgramDevirt(ExportSummary, ImportSummary);
615 }
616 
618  ModuleAnalysisManager &AM) {
619  auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
620  auto AARGetter = [&](Function &F) -> AAResults & {
621  return FAM.getResult<AAManager>(F);
622  };
623  auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & {
624  return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F);
625  };
626  auto LookupDomTree = [&FAM](Function &F) -> DominatorTree & {
627  return FAM.getResult<DominatorTreeAnalysis>(F);
628  };
629  if (!DevirtModule(M, AARGetter, OREGetter, LookupDomTree, ExportSummary,
630  ImportSummary)
631  .run())
632  return PreservedAnalyses::all();
633  return PreservedAnalyses::none();
634 }
635 
636 bool DevirtModule::runForTesting(
637  Module &M, function_ref<AAResults &(Function &)> AARGetter,
639  function_ref<DominatorTree &(Function &)> LookupDomTree) {
640  ModuleSummaryIndex Summary(/*HaveGVs=*/false);
641 
642  // Handle the command-line summary arguments. This code is for testing
643  // purposes only, so we handle errors directly.
644  if (!ClReadSummary.empty()) {
645  ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary +
646  ": ");
647  auto ReadSummaryFile =
649 
650  yaml::Input In(ReadSummaryFile->getBuffer());
651  In >> Summary;
652  ExitOnErr(errorCodeToError(In.error()));
653  }
654 
655  bool Changed =
656  DevirtModule(
657  M, AARGetter, OREGetter, LookupDomTree,
658  ClSummaryAction == PassSummaryAction::Export ? &Summary : nullptr,
659  ClSummaryAction == PassSummaryAction::Import ? &Summary : nullptr)
660  .run();
661 
662  if (!ClWriteSummary.empty()) {
663  ExitOnError ExitOnErr(
664  "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": ");
665  std::error_code EC;
667  ExitOnErr(errorCodeToError(EC));
668 
669  yaml::Output Out(OS);
670  Out << Summary;
671  }
672 
673  return Changed;
674 }
675 
676 void DevirtModule::buildTypeIdentifierMap(
677  std::vector<VTableBits> &Bits,
678  DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) {
680  Bits.reserve(M.getGlobalList().size());
682  for (GlobalVariable &GV : M.globals()) {
683  Types.clear();
684  GV.getMetadata(LLVMContext::MD_type, Types);
685  if (GV.isDeclaration() || Types.empty())
686  continue;
687 
688  VTableBits *&BitsPtr = GVToBits[&GV];
689  if (!BitsPtr) {
690  Bits.emplace_back();
691  Bits.back().GV = &GV;
692  Bits.back().ObjectSize =
693  M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType());
694  BitsPtr = &Bits.back();
695  }
696 
697  for (MDNode *Type : Types) {
698  auto TypeID = Type->getOperand(1).get();
699 
700  uint64_t Offset =
701  cast<ConstantInt>(
702  cast<ConstantAsMetadata>(Type->getOperand(0))->getValue())
703  ->getZExtValue();
704 
705  TypeIdMap[TypeID].insert({BitsPtr, Offset});
706  }
707  }
708 }
709 
710 Constant *DevirtModule::getPointerAtOffset(Constant *I, uint64_t Offset) {
711  if (I->getType()->isPointerTy()) {
712  if (Offset == 0)
713  return I;
714  return nullptr;
715  }
716 
717  const DataLayout &DL = M.getDataLayout();
718 
719  if (auto *C = dyn_cast<ConstantStruct>(I)) {
720  const StructLayout *SL = DL.getStructLayout(C->getType());
721  if (Offset >= SL->getSizeInBytes())
722  return nullptr;
723 
724  unsigned Op = SL->getElementContainingOffset(Offset);
725  return getPointerAtOffset(cast<Constant>(I->getOperand(Op)),
726  Offset - SL->getElementOffset(Op));
727  }
728  if (auto *C = dyn_cast<ConstantArray>(I)) {
729  ArrayType *VTableTy = C->getType();
730  uint64_t ElemSize = DL.getTypeAllocSize(VTableTy->getElementType());
731 
732  unsigned Op = Offset / ElemSize;
733  if (Op >= C->getNumOperands())
734  return nullptr;
735 
736  return getPointerAtOffset(cast<Constant>(I->getOperand(Op)),
737  Offset % ElemSize);
738  }
739  return nullptr;
740 }
741 
742 bool DevirtModule::tryFindVirtualCallTargets(
743  std::vector<VirtualCallTarget> &TargetsForSlot,
744  const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset) {
745  for (const TypeMemberInfo &TM : TypeMemberInfos) {
746  if (!TM.Bits->GV->isConstant())
747  return false;
748 
749  Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(),
750  TM.Offset + ByteOffset);
751  if (!Ptr)
752  return false;
753 
754  auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts());
755  if (!Fn)
756  return false;
757 
758  // We can disregard __cxa_pure_virtual as a possible call target, as
759  // calls to pure virtuals are UB.
760  if (Fn->getName() == "__cxa_pure_virtual")
761  continue;
762 
763  TargetsForSlot.push_back({Fn, &TM});
764  }
765 
766  // Give up if we couldn't find any targets.
767  return !TargetsForSlot.empty();
768 }
769 
770 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo,
771  Constant *TheFn, bool &IsExported) {
772  auto Apply = [&](CallSiteInfo &CSInfo) {
773  for (auto &&VCallSite : CSInfo.CallSites) {
774  if (RemarksEnabled)
775  VCallSite.emitRemark("single-impl",
776  TheFn->stripPointerCasts()->getName(), OREGetter);
777  VCallSite.CS.setCalledFunction(ConstantExpr::getBitCast(
778  TheFn, VCallSite.CS.getCalledValue()->getType()));
779  // This use is no longer unsafe.
780  if (VCallSite.NumUnsafeUses)
781  --*VCallSite.NumUnsafeUses;
782  }
783  if (CSInfo.isExported())
784  IsExported = true;
785  CSInfo.markDevirt();
786  };
787  Apply(SlotInfo.CSInfo);
788  for (auto &P : SlotInfo.ConstCSInfo)
789  Apply(P.second);
790 }
791 
792 bool DevirtModule::trySingleImplDevirt(
793  MutableArrayRef<VirtualCallTarget> TargetsForSlot,
794  VTableSlotInfo &SlotInfo, WholeProgramDevirtResolution *Res) {
795  // See if the program contains a single implementation of this virtual
796  // function.
797  Function *TheFn = TargetsForSlot[0].Fn;
798  for (auto &&Target : TargetsForSlot)
799  if (TheFn != Target.Fn)
800  return false;
801 
802  // If so, update each call site to call that implementation directly.
803  if (RemarksEnabled)
804  TargetsForSlot[0].WasDevirt = true;
805 
806  bool IsExported = false;
807  applySingleImplDevirt(SlotInfo, TheFn, IsExported);
808  if (!IsExported)
809  return false;
810 
811  // If the only implementation has local linkage, we must promote to external
812  // to make it visible to thin LTO objects. We can only get here during the
813  // ThinLTO export phase.
814  if (TheFn->hasLocalLinkage()) {
815  std::string NewName = (TheFn->getName() + "$merged").str();
816 
817  // Since we are renaming the function, any comdats with the same name must
818  // also be renamed. This is required when targeting COFF, as the comdat name
819  // must match one of the names of the symbols in the comdat.
820  if (Comdat *C = TheFn->getComdat()) {
821  if (C->getName() == TheFn->getName()) {
822  Comdat *NewC = M.getOrInsertComdat(NewName);
823  NewC->setSelectionKind(C->getSelectionKind());
824  for (GlobalObject &GO : M.global_objects())
825  if (GO.getComdat() == C)
826  GO.setComdat(NewC);
827  }
828  }
829 
832  TheFn->setName(NewName);
833  }
834 
836  Res->SingleImplName = TheFn->getName();
837 
838  return true;
839 }
840 
841 void DevirtModule::tryICallBranchFunnel(
842  MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo,
843  WholeProgramDevirtResolution *Res, VTableSlot Slot) {
844  Triple T(M.getTargetTriple());
845  if (T.getArch() != Triple::x86_64)
846  return;
847 
848  if (TargetsForSlot.size() > ClThreshold)
849  return;
850 
851  bool HasNonDevirt = !SlotInfo.CSInfo.AllCallSitesDevirted;
852  if (!HasNonDevirt)
853  for (auto &P : SlotInfo.ConstCSInfo)
854  if (!P.second.AllCallSitesDevirted) {
855  HasNonDevirt = true;
856  break;
857  }
858 
859  if (!HasNonDevirt)
860  return;
861 
862  FunctionType *FT =
863  FunctionType::get(Type::getVoidTy(M.getContext()), {Int8PtrTy}, true);
864  Function *JT;
865  if (isa<MDString>(Slot.TypeID)) {
867  getGlobalName(Slot, {}, "branch_funnel"), &M);
868  JT->setVisibility(GlobalValue::HiddenVisibility);
869  } else {
870  JT = Function::Create(FT, Function::InternalLinkage, "branch_funnel", &M);
871  }
872  JT->addAttribute(1, Attribute::Nest);
873 
874  std::vector<Value *> JTArgs;
875  JTArgs.push_back(JT->arg_begin());
876  for (auto &T : TargetsForSlot) {
877  JTArgs.push_back(getMemberAddr(T.TM));
878  JTArgs.push_back(T.Fn);
879  }
880 
881  BasicBlock *BB = BasicBlock::Create(M.getContext(), "", JT, nullptr);
882  Constant *Intr =
883  Intrinsic::getDeclaration(&M, llvm::Intrinsic::icall_branch_funnel, {});
884 
885  auto *CI = CallInst::Create(Intr, JTArgs, "", BB);
886  CI->setTailCallKind(CallInst::TCK_MustTail);
887  ReturnInst::Create(M.getContext(), nullptr, BB);
888 
889  bool IsExported = false;
890  applyICallBranchFunnel(SlotInfo, JT, IsExported);
891  if (IsExported)
893 }
894 
895 void DevirtModule::applyICallBranchFunnel(VTableSlotInfo &SlotInfo,
896  Constant *JT, bool &IsExported) {
897  auto Apply = [&](CallSiteInfo &CSInfo) {
898  if (CSInfo.isExported())
899  IsExported = true;
900  if (CSInfo.AllCallSitesDevirted)
901  return;
902  for (auto &&VCallSite : CSInfo.CallSites) {
903  CallSite CS = VCallSite.CS;
904 
905  // Jump tables are only profitable if the retpoline mitigation is enabled.
906  Attribute FSAttr = CS.getCaller()->getFnAttribute("target-features");
907  if (FSAttr.hasAttribute(Attribute::None) ||
908  !FSAttr.getValueAsString().contains("+retpoline"))
909  continue;
910 
911  if (RemarksEnabled)
912  VCallSite.emitRemark("branch-funnel",
913  JT->stripPointerCasts()->getName(), OREGetter);
914 
915  // Pass the address of the vtable in the nest register, which is r10 on
916  // x86_64.
917  std::vector<Type *> NewArgs;
918  NewArgs.push_back(Int8PtrTy);
919  for (Type *T : CS.getFunctionType()->params())
920  NewArgs.push_back(T);
923  CS.getFunctionType()->isVarArg()));
924 
925  IRBuilder<> IRB(CS.getInstruction());
926  std::vector<Value *> Args;
927  Args.push_back(IRB.CreateBitCast(VCallSite.VTable, Int8PtrTy));
928  for (unsigned I = 0; I != CS.getNumArgOperands(); ++I)
929  Args.push_back(CS.getArgOperand(I));
930 
931  CallSite NewCS;
932  if (CS.isCall())
933  NewCS = IRB.CreateCall(IRB.CreateBitCast(JT, NewFT), Args);
934  else
935  NewCS = IRB.CreateInvoke(
936  IRB.CreateBitCast(JT, NewFT),
937  cast<InvokeInst>(CS.getInstruction())->getNormalDest(),
938  cast<InvokeInst>(CS.getInstruction())->getUnwindDest(), Args);
939  NewCS.setCallingConv(CS.getCallingConv());
940 
942  std::vector<AttributeSet> NewArgAttrs;
943  NewArgAttrs.push_back(AttributeSet::get(
945  M.getContext(), Attribute::Nest)}));
946  for (unsigned I = 0; I + 2 < Attrs.getNumAttrSets(); ++I)
947  NewArgAttrs.push_back(Attrs.getParamAttributes(I));
948  NewCS.setAttributes(
950  Attrs.getRetAttributes(), NewArgAttrs));
951 
952  CS->replaceAllUsesWith(NewCS.getInstruction());
953  CS->eraseFromParent();
954 
955  // This use is no longer unsafe.
956  if (VCallSite.NumUnsafeUses)
957  --*VCallSite.NumUnsafeUses;
958  }
959  // Don't mark as devirtualized because there may be callers compiled without
960  // retpoline mitigation, which would mean that they are lowered to
961  // llvm.type.test and therefore require an llvm.type.test resolution for the
962  // type identifier.
963  };
964  Apply(SlotInfo.CSInfo);
965  for (auto &P : SlotInfo.ConstCSInfo)
966  Apply(P.second);
967 }
968 
969 bool DevirtModule::tryEvaluateFunctionsWithArgs(
970  MutableArrayRef<VirtualCallTarget> TargetsForSlot,
972  // Evaluate each function and store the result in each target's RetVal
973  // field.
974  for (VirtualCallTarget &Target : TargetsForSlot) {
975  if (Target.Fn->arg_size() != Args.size() + 1)
976  return false;
977 
978  Evaluator Eval(M.getDataLayout(), nullptr);
980  EvalArgs.push_back(
981  Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0)));
982  for (unsigned I = 0; I != Args.size(); ++I) {
983  auto *ArgTy = dyn_cast<IntegerType>(
984  Target.Fn->getFunctionType()->getParamType(I + 1));
985  if (!ArgTy)
986  return false;
987  EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I]));
988  }
989 
990  Constant *RetVal;
991  if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) ||
992  !isa<ConstantInt>(RetVal))
993  return false;
994  Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue();
995  }
996  return true;
997 }
998 
999 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
1000  uint64_t TheRetVal) {
1001  for (auto Call : CSInfo.CallSites)
1002  Call.replaceAndErase(
1003  "uniform-ret-val", FnName, RemarksEnabled, OREGetter,
1004  ConstantInt::get(cast<IntegerType>(Call.CS.getType()), TheRetVal));
1005  CSInfo.markDevirt();
1006 }
1007 
1008 bool DevirtModule::tryUniformRetValOpt(
1009  MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo,
1011  // Uniform return value optimization. If all functions return the same
1012  // constant, replace all calls with that constant.
1013  uint64_t TheRetVal = TargetsForSlot[0].RetVal;
1014  for (const VirtualCallTarget &Target : TargetsForSlot)
1015  if (Target.RetVal != TheRetVal)
1016  return false;
1017 
1018  if (CSInfo.isExported()) {
1020  Res->Info = TheRetVal;
1021  }
1022 
1023  applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal);
1024  if (RemarksEnabled)
1025  for (auto &&Target : TargetsForSlot)
1026  Target.WasDevirt = true;
1027  return true;
1028 }
1029 
1030 std::string DevirtModule::getGlobalName(VTableSlot Slot,
1031  ArrayRef<uint64_t> Args,
1032  StringRef Name) {
1033  std::string FullName = "__typeid_";
1034  raw_string_ostream OS(FullName);
1035  OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset;
1036  for (uint64_t Arg : Args)
1037  OS << '_' << Arg;
1038  OS << '_' << Name;
1039  return OS.str();
1040 }
1041 
1042 bool DevirtModule::shouldExportConstantsAsAbsoluteSymbols() {
1043  Triple T(M.getTargetTriple());
1044  return (T.getArch() == Triple::x86 || T.getArch() == Triple::x86_64) &&
1045  T.getObjectFormat() == Triple::ELF;
1046 }
1047 
1048 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
1049  StringRef Name, Constant *C) {
1051  getGlobalName(Slot, Args, Name), C, &M);
1053 }
1054 
1055 void DevirtModule::exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
1056  StringRef Name, uint32_t Const,
1057  uint32_t &Storage) {
1058  if (shouldExportConstantsAsAbsoluteSymbols()) {
1059  exportGlobal(
1060  Slot, Args, Name,
1061  ConstantExpr::getIntToPtr(ConstantInt::get(Int32Ty, Const), Int8PtrTy));
1062  return;
1063  }
1064 
1065  Storage = Const;
1066 }
1067 
1068 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
1069  StringRef Name) {
1070  Constant *C = M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Ty);
1071  auto *GV = dyn_cast<GlobalVariable>(C);
1072  if (GV)
1074  return C;
1075 }
1076 
1077 Constant *DevirtModule::importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
1078  StringRef Name, IntegerType *IntTy,
1079  uint32_t Storage) {
1080  if (!shouldExportConstantsAsAbsoluteSymbols())
1081  return ConstantInt::get(IntTy, Storage);
1082 
1083  Constant *C = importGlobal(Slot, Args, Name);
1084  auto *GV = cast<GlobalVariable>(C->stripPointerCasts());
1085  C = ConstantExpr::getPtrToInt(C, IntTy);
1086 
1087  // We only need to set metadata if the global is newly created, in which
1088  // case it would not have hidden visibility.
1089  if (GV->hasMetadata(LLVMContext::MD_absolute_symbol))
1090  return C;
1091 
1092  auto SetAbsRange = [&](uint64_t Min, uint64_t Max) {
1093  auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min));
1094  auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max));
1095  GV->setMetadata(LLVMContext::MD_absolute_symbol,
1096  MDNode::get(M.getContext(), {MinC, MaxC}));
1097  };
1098  unsigned AbsWidth = IntTy->getBitWidth();
1099  if (AbsWidth == IntPtrTy->getBitWidth())
1100  SetAbsRange(~0ull, ~0ull); // Full set.
1101  else
1102  SetAbsRange(0, 1ull << AbsWidth);
1103  return C;
1104 }
1105 
1106 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
1107  bool IsOne,
1108  Constant *UniqueMemberAddr) {
1109  for (auto &&Call : CSInfo.CallSites) {
1110  IRBuilder<> B(Call.CS.getInstruction());
1111  Value *Cmp =
1112  B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
1113  B.CreateBitCast(Call.VTable, Int8PtrTy), UniqueMemberAddr);
1114  Cmp = B.CreateZExt(Cmp, Call.CS->getType());
1115  Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter,
1116  Cmp);
1117  }
1118  CSInfo.markDevirt();
1119 }
1120 
1121 Constant *DevirtModule::getMemberAddr(const TypeMemberInfo *M) {
1122  Constant *C = ConstantExpr::getBitCast(M->Bits->GV, Int8PtrTy);
1123  return ConstantExpr::getGetElementPtr(Int8Ty, C,
1124  ConstantInt::get(Int64Ty, M->Offset));
1125 }
1126 
1127 bool DevirtModule::tryUniqueRetValOpt(
1128  unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot,
1129  CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res,
1130  VTableSlot Slot, ArrayRef<uint64_t> Args) {
1131  // IsOne controls whether we look for a 0 or a 1.
1132  auto tryUniqueRetValOptFor = [&](bool IsOne) {
1133  const TypeMemberInfo *UniqueMember = nullptr;
1134  for (const VirtualCallTarget &Target : TargetsForSlot) {
1135  if (Target.RetVal == (IsOne ? 1 : 0)) {
1136  if (UniqueMember)
1137  return false;
1138  UniqueMember = Target.TM;
1139  }
1140  }
1141 
1142  // We should have found a unique member or bailed out by now. We already
1143  // checked for a uniform return value in tryUniformRetValOpt.
1144  assert(UniqueMember);
1145 
1146  Constant *UniqueMemberAddr = getMemberAddr(UniqueMember);
1147  if (CSInfo.isExported()) {
1149  Res->Info = IsOne;
1150 
1151  exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr);
1152  }
1153 
1154  // Replace each call with the comparison.
1155  applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne,
1156  UniqueMemberAddr);
1157 
1158  // Update devirtualization statistics for targets.
1159  if (RemarksEnabled)
1160  for (auto &&Target : TargetsForSlot)
1161  Target.WasDevirt = true;
1162 
1163  return true;
1164  };
1165 
1166  if (BitWidth == 1) {
1167  if (tryUniqueRetValOptFor(true))
1168  return true;
1169  if (tryUniqueRetValOptFor(false))
1170  return true;
1171  }
1172  return false;
1173 }
1174 
1175 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName,
1176  Constant *Byte, Constant *Bit) {
1177  for (auto Call : CSInfo.CallSites) {
1178  auto *RetType = cast<IntegerType>(Call.CS.getType());
1179  IRBuilder<> B(Call.CS.getInstruction());
1180  Value *Addr =
1181  B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte);
1182  if (RetType->getBitWidth() == 1) {
1183  Value *Bits = B.CreateLoad(Addr);
1184  Value *BitsAndBit = B.CreateAnd(Bits, Bit);
1185  auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0));
1186  Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled,
1187  OREGetter, IsBitSet);
1188  } else {
1189  Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo());
1190  Value *Val = B.CreateLoad(RetType, ValAddr);
1191  Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled,
1192  OREGetter, Val);
1193  }
1194  }
1195  CSInfo.markDevirt();
1196 }
1197 
1198 bool DevirtModule::tryVirtualConstProp(
1199  MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo,
1200  WholeProgramDevirtResolution *Res, VTableSlot Slot) {
1201  // This only works if the function returns an integer.
1202  auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType());
1203  if (!RetType)
1204  return false;
1205  unsigned BitWidth = RetType->getBitWidth();
1206  if (BitWidth > 64)
1207  return false;
1208 
1209  // Make sure that each function is defined, does not access memory, takes at
1210  // least one argument, does not use its first argument (which we assume is
1211  // 'this'), and has the same return type.
1212  //
1213  // Note that we test whether this copy of the function is readnone, rather
1214  // than testing function attributes, which must hold for any copy of the
1215  // function, even a less optimized version substituted at link time. This is
1216  // sound because the virtual constant propagation optimizations effectively
1217  // inline all implementations of the virtual function into each call site,
1218  // rather than using function attributes to perform local optimization.
1219  for (VirtualCallTarget &Target : TargetsForSlot) {
1220  if (Target.Fn->isDeclaration() ||
1221  computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) !=
1222  MAK_ReadNone ||
1223  Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() ||
1224  Target.Fn->getReturnType() != RetType)
1225  return false;
1226  }
1227 
1228  for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) {
1229  if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first))
1230  continue;
1231 
1232  WholeProgramDevirtResolution::ByArg *ResByArg = nullptr;
1233  if (Res)
1234  ResByArg = &Res->ResByArg[CSByConstantArg.first];
1235 
1236  if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg))
1237  continue;
1238 
1239  if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second,
1240  ResByArg, Slot, CSByConstantArg.first))
1241  continue;
1242 
1243  // Find an allocation offset in bits in all vtables associated with the
1244  // type.
1245  uint64_t AllocBefore =
1246  findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth);
1247  uint64_t AllocAfter =
1248  findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth);
1249 
1250  // Calculate the total amount of padding needed to store a value at both
1251  // ends of the object.
1252  uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0;
1253  for (auto &&Target : TargetsForSlot) {
1254  TotalPaddingBefore += std::max<int64_t>(
1255  (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0);
1256  TotalPaddingAfter += std::max<int64_t>(
1257  (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0);
1258  }
1259 
1260  // If the amount of padding is too large, give up.
1261  // FIXME: do something smarter here.
1262  if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128)
1263  continue;
1264 
1265  // Calculate the offset to the value as a (possibly negative) byte offset
1266  // and (if applicable) a bit offset, and store the values in the targets.
1267  int64_t OffsetByte;
1268  uint64_t OffsetBit;
1269  if (TotalPaddingBefore <= TotalPaddingAfter)
1270  setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte,
1271  OffsetBit);
1272  else
1273  setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte,
1274  OffsetBit);
1275 
1276  if (RemarksEnabled)
1277  for (auto &&Target : TargetsForSlot)
1278  Target.WasDevirt = true;
1279 
1280 
1281  if (CSByConstantArg.second.isExported()) {
1283  exportConstant(Slot, CSByConstantArg.first, "byte", OffsetByte,
1284  ResByArg->Byte);
1285  exportConstant(Slot, CSByConstantArg.first, "bit", 1ULL << OffsetBit,
1286  ResByArg->Bit);
1287  }
1288 
1289  // Rewrite each call to a load from OffsetByte/OffsetBit.
1290  Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte);
1291  Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit);
1292  applyVirtualConstProp(CSByConstantArg.second,
1293  TargetsForSlot[0].Fn->getName(), ByteConst, BitConst);
1294  }
1295  return true;
1296 }
1297 
1298 void DevirtModule::rebuildGlobal(VTableBits &B) {
1299  if (B.Before.Bytes.empty() && B.After.Bytes.empty())
1300  return;
1301 
1302  // Align each byte array to pointer width.
1303  unsigned PointerSize = M.getDataLayout().getPointerSize();
1304  B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), PointerSize));
1305  B.After.Bytes.resize(alignTo(B.After.Bytes.size(), PointerSize));
1306 
1307  // Before was stored in reverse order; flip it now.
1308  for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I)
1309  std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]);
1310 
1311  // Build an anonymous global containing the before bytes, followed by the
1312  // original initializer, followed by the after bytes.
1313  auto NewInit = ConstantStruct::getAnon(
1314  {ConstantDataArray::get(M.getContext(), B.Before.Bytes),
1315  B.GV->getInitializer(),
1316  ConstantDataArray::get(M.getContext(), B.After.Bytes)});
1317  auto NewGV =
1318  new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(),
1319  GlobalVariable::PrivateLinkage, NewInit, "", B.GV);
1320  NewGV->setSection(B.GV->getSection());
1321  NewGV->setComdat(B.GV->getComdat());
1322 
1323  // Copy the original vtable's metadata to the anonymous global, adjusting
1324  // offsets as required.
1325  NewGV->copyMetadata(B.GV, B.Before.Bytes.size());
1326 
1327  // Build an alias named after the original global, pointing at the second
1328  // element (the original initializer).
1329  auto Alias = GlobalAlias::create(
1330  B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "",
1332  NewInit->getType(), NewGV,
1334  ConstantInt::get(Int32Ty, 1)}),
1335  &M);
1336  Alias->setVisibility(B.GV->getVisibility());
1337  Alias->takeName(B.GV);
1338 
1339  B.GV->replaceAllUsesWith(Alias);
1340  B.GV->eraseFromParent();
1341 }
1342 
1343 bool DevirtModule::areRemarksEnabled() {
1344  const auto &FL = M.getFunctionList();
1345  for (const Function &Fn : FL) {
1346  const auto &BBL = Fn.getBasicBlockList();
1347  if (BBL.empty())
1348  continue;
1349  auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front());
1350  return DI.isEnabled();
1351  }
1352  return false;
1353 }
1354 
1355 void DevirtModule::scanTypeTestUsers(Function *TypeTestFunc,
1356  Function *AssumeFunc) {
1357  // Find all virtual calls via a virtual table pointer %p under an assumption
1358  // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p
1359  // points to a member of the type identifier %md. Group calls by (type ID,
1360  // offset) pair (effectively the identity of the virtual function) and store
1361  // to CallSlots.
1362  DenseSet<CallSite> SeenCallSites;
1363  for (auto I = TypeTestFunc->use_begin(), E = TypeTestFunc->use_end();
1364  I != E;) {
1365  auto CI = dyn_cast<CallInst>(I->getUser());
1366  ++I;
1367  if (!CI)
1368  continue;
1369 
1370  // Search for virtual calls based on %p and add them to DevirtCalls.
1371  SmallVector<DevirtCallSite, 1> DevirtCalls;
1373  auto &DT = LookupDomTree(*CI->getFunction());
1374  findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT);
1375 
1376  // If we found any, add them to CallSlots.
1377  if (!Assumes.empty()) {
1378  Metadata *TypeId =
1379  cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata();
1380  Value *Ptr = CI->getArgOperand(0)->stripPointerCasts();
1381  for (DevirtCallSite Call : DevirtCalls) {
1382  // Only add this CallSite if we haven't seen it before. The vtable
1383  // pointer may have been CSE'd with pointers from other call sites,
1384  // and we don't want to process call sites multiple times. We can't
1385  // just skip the vtable Ptr if it has been seen before, however, since
1386  // it may be shared by type tests that dominate different calls.
1387  if (SeenCallSites.insert(Call.CS).second)
1388  CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS, nullptr);
1389  }
1390  }
1391 
1392  // We no longer need the assumes or the type test.
1393  for (auto Assume : Assumes)
1394  Assume->eraseFromParent();
1395  // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we
1396  // may use the vtable argument later.
1397  if (CI->use_empty())
1398  CI->eraseFromParent();
1399  }
1400 }
1401 
1402 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) {
1403  Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test);
1404 
1405  for (auto I = TypeCheckedLoadFunc->use_begin(),
1406  E = TypeCheckedLoadFunc->use_end();
1407  I != E;) {
1408  auto CI = dyn_cast<CallInst>(I->getUser());
1409  ++I;
1410  if (!CI)
1411  continue;
1412 
1413  Value *Ptr = CI->getArgOperand(0);
1414  Value *Offset = CI->getArgOperand(1);
1415  Value *TypeIdValue = CI->getArgOperand(2);
1416  Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata();
1417 
1418  SmallVector<DevirtCallSite, 1> DevirtCalls;
1419  SmallVector<Instruction *, 1> LoadedPtrs;
1421  bool HasNonCallUses = false;
1422  auto &DT = LookupDomTree(*CI->getFunction());
1423  findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds,
1424  HasNonCallUses, CI, DT);
1425 
1426  // Start by generating "pessimistic" code that explicitly loads the function
1427  // pointer from the vtable and performs the type check. If possible, we will
1428  // eliminate the load and the type check later.
1429 
1430  // If possible, only generate the load at the point where it is used.
1431  // This helps avoid unnecessary spills.
1432  IRBuilder<> LoadB(
1433  (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI);
1434  Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset);
1435  Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy));
1436  Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr);
1437 
1438  for (Instruction *LoadedPtr : LoadedPtrs) {
1439  LoadedPtr->replaceAllUsesWith(LoadedValue);
1440  LoadedPtr->eraseFromParent();
1441  }
1442 
1443  // Likewise for the type test.
1444  IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI);
1445  CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue});
1446 
1447  for (Instruction *Pred : Preds) {
1448  Pred->replaceAllUsesWith(TypeTestCall);
1449  Pred->eraseFromParent();
1450  }
1451 
1452  // We have already erased any extractvalue instructions that refer to the
1453  // intrinsic call, but the intrinsic may have other non-extractvalue uses
1454  // (although this is unlikely). In that case, explicitly build a pair and
1455  // RAUW it.
1456  if (!CI->use_empty()) {
1457  Value *Pair = UndefValue::get(CI->getType());
1458  IRBuilder<> B(CI);
1459  Pair = B.CreateInsertValue(Pair, LoadedValue, {0});
1460  Pair = B.CreateInsertValue(Pair, TypeTestCall, {1});
1461  CI->replaceAllUsesWith(Pair);
1462  }
1463 
1464  // The number of unsafe uses is initially the number of uses.
1465  auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall];
1466  NumUnsafeUses = DevirtCalls.size();
1467 
1468  // If the function pointer has a non-call user, we cannot eliminate the type
1469  // check, as one of those users may eventually call the pointer. Increment
1470  // the unsafe use count to make sure it cannot reach zero.
1471  if (HasNonCallUses)
1472  ++NumUnsafeUses;
1473  for (DevirtCallSite Call : DevirtCalls) {
1474  CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS,
1475  &NumUnsafeUses);
1476  }
1477 
1478  CI->eraseFromParent();
1479  }
1480 }
1481 
1482 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) {
1483  const TypeIdSummary *TidSummary =
1484  ImportSummary->getTypeIdSummary(cast<MDString>(Slot.TypeID)->getString());
1485  if (!TidSummary)
1486  return;
1487  auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset);
1488  if (ResI == TidSummary->WPDRes.end())
1489  return;
1490  const WholeProgramDevirtResolution &Res = ResI->second;
1491 
1493  // The type of the function in the declaration is irrelevant because every
1494  // call site will cast it to the correct type.
1495  auto *SingleImpl = M.getOrInsertFunction(
1496  Res.SingleImplName, Type::getVoidTy(M.getContext()));
1497 
1498  // This is the import phase so we should not be exporting anything.
1499  bool IsExported = false;
1500  applySingleImplDevirt(SlotInfo, SingleImpl, IsExported);
1501  assert(!IsExported);
1502  }
1503 
1504  for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) {
1505  auto I = Res.ResByArg.find(CSByConstantArg.first);
1506  if (I == Res.ResByArg.end())
1507  continue;
1508  auto &ResByArg = I->second;
1509  // FIXME: We should figure out what to do about the "function name" argument
1510  // to the apply* functions, as the function names are unavailable during the
1511  // importing phase. For now we just pass the empty string. This does not
1512  // impact correctness because the function names are just used for remarks.
1513  switch (ResByArg.TheKind) {
1515  applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info);
1516  break;
1518  Constant *UniqueMemberAddr =
1519  importGlobal(Slot, CSByConstantArg.first, "unique_member");
1520  applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info,
1521  UniqueMemberAddr);
1522  break;
1523  }
1525  Constant *Byte = importConstant(Slot, CSByConstantArg.first, "byte",
1526  Int32Ty, ResByArg.Byte);
1527  Constant *Bit = importConstant(Slot, CSByConstantArg.first, "bit", Int8Ty,
1528  ResByArg.Bit);
1529  applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit);
1530  break;
1531  }
1532  default:
1533  break;
1534  }
1535  }
1536 
1538  auto *JT = M.getOrInsertFunction(getGlobalName(Slot, {}, "branch_funnel"),
1539  Type::getVoidTy(M.getContext()));
1540  bool IsExported = false;
1541  applyICallBranchFunnel(SlotInfo, JT, IsExported);
1542  assert(!IsExported);
1543  }
1544 }
1545 
1546 void DevirtModule::removeRedundantTypeTests() {
1547  auto True = ConstantInt::getTrue(M.getContext());
1548  for (auto &&U : NumUnsafeUsesForTypeTest) {
1549  if (U.second == 0) {
1550  U.first->replaceAllUsesWith(True);
1551  U.first->eraseFromParent();
1552  }
1553  }
1554 }
1555 
1556 bool DevirtModule::run() {
1557  Function *TypeTestFunc =
1558  M.getFunction(Intrinsic::getName(Intrinsic::type_test));
1559  Function *TypeCheckedLoadFunc =
1560  M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load));
1561  Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume));
1562 
1563  // Normally if there are no users of the devirtualization intrinsics in the
1564  // module, this pass has nothing to do. But if we are exporting, we also need
1565  // to handle any users that appear only in the function summaries.
1566  if (!ExportSummary &&
1567  (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc ||
1568  AssumeFunc->use_empty()) &&
1569  (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty()))
1570  return false;
1571 
1572  if (TypeTestFunc && AssumeFunc)
1573  scanTypeTestUsers(TypeTestFunc, AssumeFunc);
1574 
1575  if (TypeCheckedLoadFunc)
1576  scanTypeCheckedLoadUsers(TypeCheckedLoadFunc);
1577 
1578  if (ImportSummary) {
1579  for (auto &S : CallSlots)
1580  importResolution(S.first, S.second);
1581 
1582  removeRedundantTypeTests();
1583 
1584  // The rest of the code is only necessary when exporting or during regular
1585  // LTO, so we are done.
1586  return true;
1587  }
1588 
1589  // Rebuild type metadata into a map for easy lookup.
1590  std::vector<VTableBits> Bits;
1592  buildTypeIdentifierMap(Bits, TypeIdMap);
1593  if (TypeIdMap.empty())
1594  return true;
1595 
1596  // Collect information from summary about which calls to try to devirtualize.
1597  if (ExportSummary) {
1599  for (auto &P : TypeIdMap) {
1600  if (auto *TypeId = dyn_cast<MDString>(P.first))
1601  MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back(
1602  TypeId);
1603  }
1604 
1605  for (auto &P : *ExportSummary) {
1606  for (auto &S : P.second.SummaryList) {
1607  auto *FS = dyn_cast<FunctionSummary>(S.get());
1608  if (!FS)
1609  continue;
1610  // FIXME: Only add live functions.
1611  for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) {
1612  for (Metadata *MD : MetadataByGUID[VF.GUID]) {
1613  CallSlots[{MD, VF.Offset}]
1614  .CSInfo.markSummaryHasTypeTestAssumeUsers();
1615  }
1616  }
1617  for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) {
1618  for (Metadata *MD : MetadataByGUID[VF.GUID]) {
1619  CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS);
1620  }
1621  }
1622  for (const FunctionSummary::ConstVCall &VC :
1623  FS->type_test_assume_const_vcalls()) {
1624  for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) {
1625  CallSlots[{MD, VC.VFunc.Offset}]
1626  .ConstCSInfo[VC.Args]
1627  .markSummaryHasTypeTestAssumeUsers();
1628  }
1629  }
1630  for (const FunctionSummary::ConstVCall &VC :
1631  FS->type_checked_load_const_vcalls()) {
1632  for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) {
1633  CallSlots[{MD, VC.VFunc.Offset}]
1634  .ConstCSInfo[VC.Args]
1635  .addSummaryTypeCheckedLoadUser(FS);
1636  }
1637  }
1638  }
1639  }
1640  }
1641 
1642  // For each (type, offset) pair:
1643  bool DidVirtualConstProp = false;
1644  std::map<std::string, Function*> DevirtTargets;
1645  for (auto &S : CallSlots) {
1646  // Search each of the members of the type identifier for the virtual
1647  // function implementation at offset S.first.ByteOffset, and add to
1648  // TargetsForSlot.
1649  std::vector<VirtualCallTarget> TargetsForSlot;
1650  if (tryFindVirtualCallTargets(TargetsForSlot, TypeIdMap[S.first.TypeID],
1651  S.first.ByteOffset)) {
1652  WholeProgramDevirtResolution *Res = nullptr;
1653  if (ExportSummary && isa<MDString>(S.first.TypeID))
1654  Res = &ExportSummary
1655  ->getOrInsertTypeIdSummary(
1656  cast<MDString>(S.first.TypeID)->getString())
1657  .WPDRes[S.first.ByteOffset];
1658 
1659  if (!trySingleImplDevirt(TargetsForSlot, S.second, Res)) {
1660  DidVirtualConstProp |=
1661  tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first);
1662 
1663  tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first);
1664  }
1665 
1666  // Collect functions devirtualized at least for one call site for stats.
1667  if (RemarksEnabled)
1668  for (const auto &T : TargetsForSlot)
1669  if (T.WasDevirt)
1670  DevirtTargets[T.Fn->getName()] = T.Fn;
1671  }
1672 
1673  // CFI-specific: if we are exporting and any llvm.type.checked.load
1674  // intrinsics were *not* devirtualized, we need to add the resulting
1675  // llvm.type.test intrinsics to the function summaries so that the
1676  // LowerTypeTests pass will export them.
1677  if (ExportSummary && isa<MDString>(S.first.TypeID)) {
1678  auto GUID =
1679  GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString());
1680  for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers)
1681  FS->addTypeTest(GUID);
1682  for (auto &CCS : S.second.ConstCSInfo)
1683  for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers)
1684  FS->addTypeTest(GUID);
1685  }
1686  }
1687 
1688  if (RemarksEnabled) {
1689  // Generate remarks for each devirtualized function.
1690  for (const auto &DT : DevirtTargets) {
1691  Function *F = DT.second;
1692 
1693  using namespace ore;
1694  OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F)
1695  << "devirtualized "
1696  << NV("FunctionName", F->getName()));
1697  }
1698  }
1699 
1700  removeRedundantTypeTests();
1701 
1702  // Rebuild each global we touched as part of virtual constant propagation to
1703  // include the before and after bytes.
1704  if (DidVirtualConstProp)
1705  for (VTableBits &B : Bits)
1706  rebuildGlobal(B);
1707 
1708  return true;
1709 }
void setVisibility(VisibilityTypes V)
Definition: GlobalValue.h:239
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:218
use_iterator use_end()
Definition: Value.h:347
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:239
bool empty() const
Definition: Function.h:662
bool hasLocalLinkage() const
Definition: GlobalValue.h:436
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:770
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:64
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:1150
void findDevirtualizableCallsForTypeTest(SmallVectorImpl< DevirtCallSite > &DevirtCalls, SmallVectorImpl< CallInst *> &Assumes, const CallInst *CI, DominatorTree &DT)
Given a call to the intrinsic @llvm.type.test, find all devirtualizable call sites based on the call ...
VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM)
Implements a dense probed hash-table based set.
Definition: DenseSet.h:250
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:1223
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:107
static Constant * getIntToPtr(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1737
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:864
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:231
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:685
const GlobalListType & getGlobalList() const
Get the Module&#39;s list of global variables (constant).
Definition: Module.h:511
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:268
Export information to summary.
const char * getName() const
getName - Get the target name.
iterator_range< global_object_iterator > global_objects()
Definition: Module.h:649
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 findDevirtualizableCallsForTypeCheckedLoad(SmallVectorImpl< DevirtCallSite > &DevirtCalls, SmallVectorImpl< Instruction *> &LoadedPtrs, SmallVectorImpl< Instruction *> &Preds, bool &HasNonCallUses, const CallInst *CI, DominatorTree &DT)
Given a call to the intrinsic @llvm.type.checked.load, find all devirtualizable call sites based on t...
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
amdgpu Simplify well known AMD library false Value Value const Twine & Name
StringRef getName(ID id)
Return the LLVM name for an intrinsic, such as "llvm.ppc.altivec.lvx".
Definition: Function.cpp:627
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:364
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:243
The returned value is undefined.
Definition: MathExtras.h:46
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:743
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:72
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:286
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:1641
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:451
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:430
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:410
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:145
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:1021
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:157
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:1750
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:1317
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:742
static MDTuple * get(LLVMContext &Context, ArrayRef< Metadata *> MDs)
Definition: Metadata.h:1166
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:419
unsigned getNumOperands() const
bool hasAttribute(AttrKind Val) const
Return true if the attribute is present.
Definition: Attributes.cpp:202
static Function * Create(FunctionType *Ty, LinkageTypes Linkage, unsigned AddrSpace, const Twine &N="", Module *M=nullptr)
Definition: Function.h:136
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:154
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
static AttributeSet get(LLVMContext &C, const AttrBuilder &B)
Definition: Attributes.cpp:513
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:120
VisibilityTypes getVisibility() const
Definition: GlobalValue.h:233
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:58
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:205
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:87
ValuesClass values(OptsTy... Options)
Helper to build a ValuesClass by forwarding a variable number of arguments as an initializer list to ...
Definition: CommandLine.h:639
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:224
#define DEBUG_TYPE
A manager for alias analyses.
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:359
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:1102
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:100
void reserve(size_type NumEntries)
Grow the densemap so that it can contain at least NumEntries items before resizing again...
Definition: DenseMap.h:130
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:1392
const Constant * stripPointerCasts() const
Definition: Constant.h:173
Comdat * getOrInsertComdat(StringRef Name)
Return the Comdat in the module with the specified name.
Definition: Module.cpp:477
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:530
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:160
unsigned getNumArgOperands() const
Definition: CallSite.h:293
size_t size() const
Definition: SmallVector.h:53
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:499
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:39
Value * CreateGEP(Value *Ptr, ArrayRef< Value *> IdxList, const Twine &Name="")
Definition: IRBuilder.h:1386
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
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:847
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:621
static BranchInst * Create(BasicBlock *IfTrue, Instruction *InsertBefore=nullptr)
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:577
BBTy * getParent() const
Get the basic block containing the call site.
Definition: CallSite.h:97
void setLinkage(LinkageTypes LT)
Definition: GlobalValue.h:445
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:941
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:501
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:614
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:339
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:366
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:56
StringRef getValueAsString() const
Return the attribute&#39;s value as a string.
Definition: Attributes.cpp:195
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:215
static Constant * getPtrToInt(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:1724
#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:633
uint32_t Size
Definition: Profile.cpp:47
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:81
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())
A raw_ostream that writes to an std::string.
Definition: raw_ostream.h:483
aarch64 promote const
LLVM Value Representation.
Definition: Value.h:73
FunctionType * getFunctionType() const
Definition: CallSite.h:320
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:331
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:574
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.
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:260
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:189
The optimization diagnostic interface.
bool use_empty() const
Definition: Value.h:323
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:423
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:873
IntegerType * Int32Ty
CallInst * CreateCall(Value *Callee, ArrayRef< Value *> Args=None, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1883
An analysis over an "outer" IR unit that provides access to an analysis manager over an "inner" IR un...
Definition: PassManager.h:1038