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 
44 #include "llvm/Transforms/IPO/WholeProgramDevirt.h"
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/iterator_range.h"
50 #include "llvm/ADT/MapVector.h"
51 #include "llvm/ADT/SmallVector.h"
52 #include "llvm/Analysis/AliasAnalysis.h"
53 #include "llvm/Analysis/BasicAliasAnalysis.h"
54 #include "llvm/Analysis/TypeMetadataUtils.h"
55 #include "llvm/IR/CallSite.h"
56 #include "llvm/IR/Constants.h"
57 #include "llvm/IR/DataLayout.h"
58 #include "llvm/IR/DebugInfoMetadata.h"
59 #include "llvm/IR/DebugLoc.h"
60 #include "llvm/IR/DerivedTypes.h"
61 #include "llvm/IR/DiagnosticInfo.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"
73 #include "llvm/IR/ModuleSummaryIndexYAML.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"
79 #include "llvm/Support/FileSystem.h"
80 #include "llvm/Support/MathExtras.h"
81 #include "llvm/Transforms/IPO.h"
82 #include "llvm/Transforms/IPO/FunctionAttrs.h"
83 #include "llvm/Transforms/Utils/Evaluator.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 
95 static cl::opt<PassSummaryAction> ClSummaryAction(
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"),
99                clEnumValN(PassSummaryAction::Import, "import",
100                           "Import typeid resolutions from summary and globals"),
101                clEnumValN(PassSummaryAction::Export, "export",
102                           "Export typeid resolutions to summary and globals")),
103     cl::Hidden);
104 
105 static cl::opt<std::string> ClReadSummary(
106     "wholeprogramdevirt-read-summary",
107     cl::desc("Read summary from given YAML file before running pass"),
108     cl::Hidden);
109 
110 static cl::opt<std::string> ClWriteSummary(
111     "wholeprogramdevirt-write-summary",
112     cl::desc("Write summary to given YAML file after running pass"),
113     cl::Hidden);
114 
115 // Find the minimum offset that we may store a value of size Size bits at. If
116 // IsAfter is set, look for an offset before the object, otherwise look for an
117 // offset after the object.
118 uint64_t
119 wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets,
120                                      bool IsAfter, uint64_t Size) {
121   // Find a minimum offset taking into account only vtable sizes.
122   uint64_t MinByte = 0;
123   for (const VirtualCallTarget &Target : Targets) {
124     if (IsAfter)
125       MinByte = std::max(MinByte, Target.minAfterBytes());
126     else
127       MinByte = std::max(MinByte, Target.minBeforeBytes());
128   }
129 
130   // Build a vector of arrays of bytes covering, for each target, a slice of the
131   // used region (see AccumBitVector::BytesUsed in
132   // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively,
133   // this aligns the used regions to start at MinByte.
134   //
135   // In this example, A, B and C are vtables, # is a byte already allocated for
136   // a virtual function pointer, AAAA... (etc.) are the used regions for the
137   // vtables and Offset(X) is the value computed for the Offset variable below
138   // for X.
139   //
140   //                    Offset(A)
141   //                    |       |
142   //                            |MinByte
143   // A: ################AAAAAAAA|AAAAAAAA
144   // B: ########BBBBBBBBBBBBBBBB|BBBB
145   // C: ########################|CCCCCCCCCCCCCCCC
146   //            |   Offset(B)   |
147   //
148   // This code produces the slices of A, B and C that appear after the divider
149   // at MinByte.
150   std::vector<ArrayRef<uint8_t>> Used;
151   for (const VirtualCallTarget &Target : Targets) {
152     ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed
153                                        : Target.TM->Bits->Before.BytesUsed;
154     uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes()
155                               : MinByte - Target.minBeforeBytes();
156 
157     // Disregard used regions that are smaller than Offset. These are
158     // effectively all-free regions that do not need to be checked.
159     if (VTUsed.size() > Offset)
160       Used.push_back(VTUsed.slice(Offset));
161   }
162 
163   if (Size == 1) {
164     // Find a free bit in each member of Used.
165     for (unsigned I = 0;; ++I) {
166       uint8_t BitsUsed = 0;
167       for (auto &&B : Used)
168         if (I < B.size())
169           BitsUsed |= B[I];
170       if (BitsUsed != 0xff)
171         return (MinByte + I) * 8 +
172                countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined);
173     }
174   } else {
175     // Find a free (Size/8) byte region in each member of Used.
176     // FIXME: see if alignment helps.
177     for (unsigned I = 0;; ++I) {
178       for (auto &&B : Used) {
179         unsigned Byte = 0;
180         while ((I + Byte) < B.size() && Byte < (Size / 8)) {
181           if (B[I + Byte])
182             goto NextI;
183           ++Byte;
184         }
185       }
186       return (MinByte + I) * 8;
187     NextI:;
188     }
189   }
190 }
191 
192 void wholeprogramdevirt::setBeforeReturnValues(
193     MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore,
194     unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
195   if (BitWidth == 1)
196     OffsetByte = -(AllocBefore / 8 + 1);
197   else
198     OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8);
199   OffsetBit = AllocBefore % 8;
200 
201   for (VirtualCallTarget &Target : Targets) {
202     if (BitWidth == 1)
203       Target.setBeforeBit(AllocBefore);
204     else
205       Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8);
206   }
207 }
208 
209 void wholeprogramdevirt::setAfterReturnValues(
210     MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter,
211     unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
212   if (BitWidth == 1)
213     OffsetByte = AllocAfter / 8;
214   else
215     OffsetByte = (AllocAfter + 7) / 8;
216   OffsetBit = AllocAfter % 8;
217 
218   for (VirtualCallTarget &Target : Targets) {
219     if (BitWidth == 1)
220       Target.setAfterBit(AllocAfter);
221     else
222       Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8);
223   }
224 }
225 
226 VirtualCallTarget::VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM)
227     : Fn(Fn), TM(TM),
228       IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {}
229 
230 namespace {
231 
232 // A slot in a set of virtual tables. The TypeID identifies the set of virtual
233 // tables, and the ByteOffset is the offset in bytes from the address point to
234 // the virtual function pointer.
235 struct VTableSlot {
236   Metadata *TypeID;
237   uint64_t ByteOffset;
238 };
239 
240 } // end anonymous namespace
241 
242 namespace llvm {
243 
244 template <> struct DenseMapInfo<VTableSlot> {
245   static VTableSlot getEmptyKey() {
246     return {DenseMapInfo<Metadata *>::getEmptyKey(),
247             DenseMapInfo<uint64_t>::getEmptyKey()};
248   }
249   static VTableSlot getTombstoneKey() {
250     return {DenseMapInfo<Metadata *>::getTombstoneKey(),
251             DenseMapInfo<uint64_t>::getTombstoneKey()};
252   }
253   static unsigned getHashValue(const VTableSlot &I) {
254     return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^
255            DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset);
256   }
257   static bool isEqual(const VTableSlot &LHS,
258                       const VTableSlot &RHS) {
259     return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset;
260   }
261 };
262 
263 } // end namespace llvm
264 
265 namespace {
266 
267 // A virtual call site. VTable is the loaded virtual table pointer, and CS is
268 // the indirect virtual call.
269 struct VirtualCallSite {
270   Value *VTable;
271   CallSite CS;
272 
273   // If non-null, this field points to the associated unsafe use count stored in
274   // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description
275   // of that field for details.
276   unsigned *NumUnsafeUses;
277 
278   void emitRemark(const Twine &OptName, const Twine &TargetName) {
279     Function *F = CS.getCaller();
280     emitOptimizationRemark(
281         F->getContext(), DEBUG_TYPE, *F,
282         CS.getInstruction()->getDebugLoc(),
283         OptName + ": devirtualized a call to " + TargetName);
284   }
285 
286   void replaceAndErase(const Twine &OptName, const Twine &TargetName,
287                        bool RemarksEnabled, Value *New) {
288     if (RemarksEnabled)
289       emitRemark(OptName, TargetName);
290     CS->replaceAllUsesWith(New);
291     if (auto II = dyn_cast<InvokeInst>(CS.getInstruction())) {
292       BranchInst::Create(II->getNormalDest(), CS.getInstruction());
293       II->getUnwindDest()->removePredecessor(II->getParent());
294     }
295     CS->eraseFromParent();
296     // This use is no longer unsafe.
297     if (NumUnsafeUses)
298       --*NumUnsafeUses;
299   }
300 };
301 
302 // Call site information collected for a specific VTableSlot and possibly a list
303 // of constant integer arguments. The grouping by arguments is handled by the
304 // VTableSlotInfo class.
305 struct CallSiteInfo {
306   /// The set of call sites for this slot. Used during regular LTO and the
307   /// import phase of ThinLTO (as well as the export phase of ThinLTO for any
308   /// call sites that appear in the merged module itself); in each of these
309   /// cases we are directly operating on the call sites at the IR level.
310   std::vector<VirtualCallSite> CallSites;
311 
312   // These fields are used during the export phase of ThinLTO and reflect
313   // information collected from function summaries.
314 
315   /// Whether any function summary contains an llvm.assume(llvm.type.test) for
316   /// this slot.
317   bool SummaryHasTypeTestAssumeUsers;
318 
319   /// CFI-specific: a vector containing the list of function summaries that use
320   /// the llvm.type.checked.load intrinsic and therefore will require
321   /// resolutions for llvm.type.test in order to implement CFI checks if
322   /// devirtualization was unsuccessful. If devirtualization was successful, the
323   /// pass will clear this vector by calling markDevirt(). If at the end of the
324   /// pass the vector is non-empty, we will need to add a use of llvm.type.test
325   /// to each of the function summaries in the vector.
326   std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers;
327 
328   bool isExported() const {
329     return SummaryHasTypeTestAssumeUsers ||
330            !SummaryTypeCheckedLoadUsers.empty();
331   }
332 
333   /// As explained in the comment for SummaryTypeCheckedLoadUsers.
334   void markDevirt() { SummaryTypeCheckedLoadUsers.clear(); }
335 };
336 
337 // Call site information collected for a specific VTableSlot.
338 struct VTableSlotInfo {
339   // The set of call sites which do not have all constant integer arguments
340   // (excluding "this").
341   CallSiteInfo CSInfo;
342 
343   // The set of call sites with all constant integer arguments (excluding
344   // "this"), grouped by argument list.
345   std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo;
346 
347   void addCallSite(Value *VTable, CallSite CS, unsigned *NumUnsafeUses);
348 
349 private:
350   CallSiteInfo &findCallSiteInfo(CallSite CS);
351 };
352 
353 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallSite CS) {
354   std::vector<uint64_t> Args;
355   auto *CI = dyn_cast<IntegerType>(CS.getType());
356   if (!CI || CI->getBitWidth() > 64 || CS.arg_empty())
357     return CSInfo;
358   for (auto &&Arg : make_range(CS.arg_begin() + 1, CS.arg_end())) {
359     auto *CI = dyn_cast<ConstantInt>(Arg);
360     if (!CI || CI->getBitWidth() > 64)
361       return CSInfo;
362     Args.push_back(CI->getZExtValue());
363   }
364   return ConstCSInfo[Args];
365 }
366 
367 void VTableSlotInfo::addCallSite(Value *VTable, CallSite CS,
368                                  unsigned *NumUnsafeUses) {
369   findCallSiteInfo(CS).CallSites.push_back({VTable, CS, NumUnsafeUses});
370 }
371 
372 struct DevirtModule {
373   Module &M;
374   function_ref<AAResults &(Function &)> AARGetter;
375 
376   PassSummaryAction Action;
377   ModuleSummaryIndex *Summary;
378 
379   IntegerType *Int8Ty;
380   PointerType *Int8PtrTy;
381   IntegerType *Int32Ty;
382   IntegerType *Int64Ty;
383   IntegerType *IntPtrTy;
384 
385   bool RemarksEnabled;
386 
387   MapVector<VTableSlot, VTableSlotInfo> CallSlots;
388 
389   // This map keeps track of the number of "unsafe" uses of a loaded function
390   // pointer. The key is the associated llvm.type.test intrinsic call generated
391   // by this pass. An unsafe use is one that calls the loaded function pointer
392   // directly. Every time we eliminate an unsafe use (for example, by
393   // devirtualizing it or by applying virtual constant propagation), we
394   // decrement the value stored in this map. If a value reaches zero, we can
395   // eliminate the type check by RAUWing the associated llvm.type.test call with
396   // true.
397   std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest;
398 
399   DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter,
400                PassSummaryAction Action, ModuleSummaryIndex *Summary)
401       : M(M), AARGetter(AARGetter), Action(Action), Summary(Summary),
402         Int8Ty(Type::getInt8Ty(M.getContext())),
403         Int8PtrTy(Type::getInt8PtrTy(M.getContext())),
404         Int32Ty(Type::getInt32Ty(M.getContext())),
405         Int64Ty(Type::getInt64Ty(M.getContext())),
406         IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)),
407         RemarksEnabled(areRemarksEnabled()) {}
408 
409   bool areRemarksEnabled();
410 
411   void scanTypeTestUsers(Function *TypeTestFunc, Function *AssumeFunc);
412   void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc);
413 
414   void buildTypeIdentifierMap(
415       std::vector<VTableBits> &Bits,
416       DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap);
417   Constant *getPointerAtOffset(Constant *I, uint64_t Offset);
418   bool
419   tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot,
420                             const std::set<TypeMemberInfo> &TypeMemberInfos,
421                             uint64_t ByteOffset);
422 
423   void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn,
424                              bool &IsExported);
425   bool trySingleImplDevirt(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
426                            VTableSlotInfo &SlotInfo,
427                            WholeProgramDevirtResolution *Res);
428 
429   bool tryEvaluateFunctionsWithArgs(
430       MutableArrayRef<VirtualCallTarget> TargetsForSlot,
431       ArrayRef<uint64_t> Args);
432 
433   void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
434                              uint64_t TheRetVal);
435   bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
436                            CallSiteInfo &CSInfo,
437                            WholeProgramDevirtResolution::ByArg *Res);
438 
439   // Returns the global symbol name that is used to export information about the
440   // given vtable slot and list of arguments.
441   std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args,
442                             StringRef Name);
443 
444   // This function is called during the export phase to create a symbol
445   // definition containing information about the given vtable slot and list of
446   // arguments.
447   void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name,
448                     Constant *C);
449 
450   // This function is called during the import phase to create a reference to
451   // the symbol definition created during the export phase.
452   Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
453                          StringRef Name, unsigned AbsWidth = 0);
454 
455   void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne,
456                             Constant *UniqueMemberAddr);
457   bool tryUniqueRetValOpt(unsigned BitWidth,
458                           MutableArrayRef<VirtualCallTarget> TargetsForSlot,
459                           CallSiteInfo &CSInfo,
460                           WholeProgramDevirtResolution::ByArg *Res,
461                           VTableSlot Slot, ArrayRef<uint64_t> Args);
462 
463   void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName,
464                              Constant *Byte, Constant *Bit);
465   bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
466                            VTableSlotInfo &SlotInfo,
467                            WholeProgramDevirtResolution *Res, VTableSlot Slot);
468 
469   void rebuildGlobal(VTableBits &B);
470 
471   // Apply the summary resolution for Slot to all virtual calls in SlotInfo.
472   void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo);
473 
474   // If we were able to eliminate all unsafe uses for a type checked load,
475   // eliminate the associated type tests by replacing them with true.
476   void removeRedundantTypeTests();
477 
478   bool run();
479 
480   // Lower the module using the action and summary passed as command line
481   // arguments. For testing purposes only.
482   static bool runForTesting(Module &M,
483                             function_ref<AAResults &(Function &)> AARGetter);
484 };
485 
486 struct WholeProgramDevirt : public ModulePass {
487   static char ID;
488 
489   bool UseCommandLine = false;
490 
491   PassSummaryAction Action;
492   ModuleSummaryIndex *Summary;
493 
494   WholeProgramDevirt() : ModulePass(ID), UseCommandLine(true) {
495     initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry());
496   }
497 
498   WholeProgramDevirt(PassSummaryAction Action, ModuleSummaryIndex *Summary)
499       : ModulePass(ID), Action(Action), Summary(Summary) {
500     initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry());
501   }
502 
503   bool runOnModule(Module &M) override {
504     if (skipModule(M))
505       return false;
506     if (UseCommandLine)
507       return DevirtModule::runForTesting(M, LegacyAARGetter(*this));
508     return DevirtModule(M, LegacyAARGetter(*this), Action, Summary).run();
509   }
510 
511   void getAnalysisUsage(AnalysisUsage &AU) const override {
512     AU.addRequired<AssumptionCacheTracker>();
513     AU.addRequired<TargetLibraryInfoWrapperPass>();
514   }
515 };
516 
517 } // end anonymous namespace
518 
519 INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt",
520                       "Whole program devirtualization", false, false)
521 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
522 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
523 INITIALIZE_PASS_END(WholeProgramDevirt, "wholeprogramdevirt",
524                     "Whole program devirtualization", false, false)
525 char WholeProgramDevirt::ID = 0;
526 
527 ModulePass *llvm::createWholeProgramDevirtPass(PassSummaryAction Action,
528                                                ModuleSummaryIndex *Summary) {
529   return new WholeProgramDevirt(Action, Summary);
530 }
531 
532 PreservedAnalyses WholeProgramDevirtPass::run(Module &M,
533                                               ModuleAnalysisManager &AM) {
534   auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
535   auto AARGetter = [&](Function &F) -> AAResults & {
536     return FAM.getResult<AAManager>(F);
537   };
538   if (!DevirtModule(M, AARGetter, PassSummaryAction::None, nullptr).run())
539     return PreservedAnalyses::all();
540   return PreservedAnalyses::none();
541 }
542 
543 bool DevirtModule::runForTesting(
544     Module &M, function_ref<AAResults &(Function &)> AARGetter) {
545   ModuleSummaryIndex Summary;
546 
547   // Handle the command-line summary arguments. This code is for testing
548   // purposes only, so we handle errors directly.
549   if (!ClReadSummary.empty()) {
550     ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary +
551                           ": ");
552     auto ReadSummaryFile =
553         ExitOnErr(errorOrToExpected(MemoryBuffer::getFile(ClReadSummary)));
554 
555     yaml::Input In(ReadSummaryFile->getBuffer());
556     In >> Summary;
557     ExitOnErr(errorCodeToError(In.error()));
558   }
559 
560   bool Changed = DevirtModule(M, AARGetter, ClSummaryAction, &Summary).run();
561 
562   if (!ClWriteSummary.empty()) {
563     ExitOnError ExitOnErr(
564         "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": ");
565     std::error_code EC;
566     raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::F_Text);
567     ExitOnErr(errorCodeToError(EC));
568 
569     yaml::Output Out(OS);
570     Out << Summary;
571   }
572 
573   return Changed;
574 }
575 
576 void DevirtModule::buildTypeIdentifierMap(
577     std::vector<VTableBits> &Bits,
578     DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) {
579   DenseMap<GlobalVariable *, VTableBits *> GVToBits;
580   Bits.reserve(M.getGlobalList().size());
581   SmallVector<MDNode *, 2> Types;
582   for (GlobalVariable &GV : M.globals()) {
583     Types.clear();
584     GV.getMetadata(LLVMContext::MD_type, Types);
585     if (Types.empty())
586       continue;
587 
588     VTableBits *&BitsPtr = GVToBits[&GV];
589     if (!BitsPtr) {
590       Bits.emplace_back();
591       Bits.back().GV = &GV;
592       Bits.back().ObjectSize =
593           M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType());
594       BitsPtr = &Bits.back();
595     }
596 
597     for (MDNode *Type : Types) {
598       auto TypeID = Type->getOperand(1).get();
599 
600       uint64_t Offset =
601           cast<ConstantInt>(
602               cast<ConstantAsMetadata>(Type->getOperand(0))->getValue())
603               ->getZExtValue();
604 
605       TypeIdMap[TypeID].insert({BitsPtr, Offset});
606     }
607   }
608 }
609 
610 Constant *DevirtModule::getPointerAtOffset(Constant *I, uint64_t Offset) {
611   if (I->getType()->isPointerTy()) {
612     if (Offset == 0)
613       return I;
614     return nullptr;
615   }
616 
617   const DataLayout &DL = M.getDataLayout();
618 
619   if (auto *C = dyn_cast<ConstantStruct>(I)) {
620     const StructLayout *SL = DL.getStructLayout(C->getType());
621     if (Offset >= SL->getSizeInBytes())
622       return nullptr;
623 
624     unsigned Op = SL->getElementContainingOffset(Offset);
625     return getPointerAtOffset(cast<Constant>(I->getOperand(Op)),
626                               Offset - SL->getElementOffset(Op));
627   }
628   if (auto *C = dyn_cast<ConstantArray>(I)) {
629     ArrayType *VTableTy = C->getType();
630     uint64_t ElemSize = DL.getTypeAllocSize(VTableTy->getElementType());
631 
632     unsigned Op = Offset / ElemSize;
633     if (Op >= C->getNumOperands())
634       return nullptr;
635 
636     return getPointerAtOffset(cast<Constant>(I->getOperand(Op)),
637                               Offset % ElemSize);
638   }
639   return nullptr;
640 }
641 
642 bool DevirtModule::tryFindVirtualCallTargets(
643     std::vector<VirtualCallTarget> &TargetsForSlot,
644     const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset) {
645   for (const TypeMemberInfo &TM : TypeMemberInfos) {
646     if (!TM.Bits->GV->isConstant())
647       return false;
648 
649     Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(),
650                                        TM.Offset + ByteOffset);
651     if (!Ptr)
652       return false;
653 
654     auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts());
655     if (!Fn)
656       return false;
657 
658     // We can disregard __cxa_pure_virtual as a possible call target, as
659     // calls to pure virtuals are UB.
660     if (Fn->getName() == "__cxa_pure_virtual")
661       continue;
662 
663     TargetsForSlot.push_back({Fn, &TM});
664   }
665 
666   // Give up if we couldn't find any targets.
667   return !TargetsForSlot.empty();
668 }
669 
670 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo,
671                                          Constant *TheFn, bool &IsExported) {
672   auto Apply = [&](CallSiteInfo &CSInfo) {
673     for (auto &&VCallSite : CSInfo.CallSites) {
674       if (RemarksEnabled)
675         VCallSite.emitRemark("single-impl", TheFn->getName());
676       VCallSite.CS.setCalledFunction(ConstantExpr::getBitCast(
677           TheFn, VCallSite.CS.getCalledValue()->getType()));
678       // This use is no longer unsafe.
679       if (VCallSite.NumUnsafeUses)
680         --*VCallSite.NumUnsafeUses;
681     }
682     if (CSInfo.isExported()) {
683       IsExported = true;
684       CSInfo.markDevirt();
685     }
686   };
687   Apply(SlotInfo.CSInfo);
688   for (auto &P : SlotInfo.ConstCSInfo)
689     Apply(P.second);
690 }
691 
692 bool DevirtModule::trySingleImplDevirt(
693     MutableArrayRef<VirtualCallTarget> TargetsForSlot,
694     VTableSlotInfo &SlotInfo, WholeProgramDevirtResolution *Res) {
695   // See if the program contains a single implementation of this virtual
696   // function.
697   Function *TheFn = TargetsForSlot[0].Fn;
698   for (auto &&Target : TargetsForSlot)
699     if (TheFn != Target.Fn)
700       return false;
701 
702   // If so, update each call site to call that implementation directly.
703   if (RemarksEnabled)
704     TargetsForSlot[0].WasDevirt = true;
705 
706   bool IsExported = false;
707   applySingleImplDevirt(SlotInfo, TheFn, IsExported);
708   if (!IsExported)
709     return false;
710 
711   // If the only implementation has local linkage, we must promote to external
712   // to make it visible to thin LTO objects. We can only get here during the
713   // ThinLTO export phase.
714   if (TheFn->hasLocalLinkage()) {
715     TheFn->setLinkage(GlobalValue::ExternalLinkage);
716     TheFn->setVisibility(GlobalValue::HiddenVisibility);
717     TheFn->setName(TheFn->getName() + "$merged");
718   }
719 
720   Res->TheKind = WholeProgramDevirtResolution::SingleImpl;
721   Res->SingleImplName = TheFn->getName();
722 
723   return true;
724 }
725 
726 bool DevirtModule::tryEvaluateFunctionsWithArgs(
727     MutableArrayRef<VirtualCallTarget> TargetsForSlot,
728     ArrayRef<uint64_t> Args) {
729   // Evaluate each function and store the result in each target's RetVal
730   // field.
731   for (VirtualCallTarget &Target : TargetsForSlot) {
732     if (Target.Fn->arg_size() != Args.size() + 1)
733       return false;
734 
735     Evaluator Eval(M.getDataLayout(), nullptr);
736     SmallVector<Constant *, 2> EvalArgs;
737     EvalArgs.push_back(
738         Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0)));
739     for (unsigned I = 0; I != Args.size(); ++I) {
740       auto *ArgTy = dyn_cast<IntegerType>(
741           Target.Fn->getFunctionType()->getParamType(I + 1));
742       if (!ArgTy)
743         return false;
744       EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I]));
745     }
746 
747     Constant *RetVal;
748     if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) ||
749         !isa<ConstantInt>(RetVal))
750       return false;
751     Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue();
752   }
753   return true;
754 }
755 
756 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
757                                          uint64_t TheRetVal) {
758   for (auto Call : CSInfo.CallSites)
759     Call.replaceAndErase(
760         "uniform-ret-val", FnName, RemarksEnabled,
761         ConstantInt::get(cast<IntegerType>(Call.CS.getType()), TheRetVal));
762   CSInfo.markDevirt();
763 }
764 
765 bool DevirtModule::tryUniformRetValOpt(
766     MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo,
767     WholeProgramDevirtResolution::ByArg *Res) {
768   // Uniform return value optimization. If all functions return the same
769   // constant, replace all calls with that constant.
770   uint64_t TheRetVal = TargetsForSlot[0].RetVal;
771   for (const VirtualCallTarget &Target : TargetsForSlot)
772     if (Target.RetVal != TheRetVal)
773       return false;
774 
775   if (CSInfo.isExported()) {
776     Res->TheKind = WholeProgramDevirtResolution::ByArg::UniformRetVal;
777     Res->Info = TheRetVal;
778   }
779 
780   applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal);
781   if (RemarksEnabled)
782     for (auto &&Target : TargetsForSlot)
783       Target.WasDevirt = true;
784   return true;
785 }
786 
787 std::string DevirtModule::getGlobalName(VTableSlot Slot,
788                                         ArrayRef<uint64_t> Args,
789                                         StringRef Name) {
790   std::string FullName = "__typeid_";
791   raw_string_ostream OS(FullName);
792   OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset;
793   for (uint64_t Arg : Args)
794     OS << '_' << Arg;
795   OS << '_' << Name;
796   return OS.str();
797 }
798 
799 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
800                                 StringRef Name, Constant *C) {
801   GlobalAlias *GA = GlobalAlias::create(Int8Ty, 0, GlobalValue::ExternalLinkage,
802                                         getGlobalName(Slot, Args, Name), C, &M);
803   GA->setVisibility(GlobalValue::HiddenVisibility);
804 }
805 
806 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
807                                      StringRef Name, unsigned AbsWidth) {
808   Constant *C = M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Ty);
809   auto *GV = dyn_cast<GlobalVariable>(C);
810   // We only need to set metadata if the global is newly created, in which
811   // case it would not have hidden visibility.
812   if (!GV || GV->getVisibility() == GlobalValue::HiddenVisibility)
813     return C;
814 
815   GV->setVisibility(GlobalValue::HiddenVisibility);
816   auto SetAbsRange = [&](uint64_t Min, uint64_t Max) {
817     auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min));
818     auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max));
819     GV->setMetadata(LLVMContext::MD_absolute_symbol,
820                     MDNode::get(M.getContext(), {MinC, MaxC}));
821   };
822   if (AbsWidth == IntPtrTy->getBitWidth())
823     SetAbsRange(~0ull, ~0ull); // Full set.
824   else if (AbsWidth)
825     SetAbsRange(0, 1ull << AbsWidth);
826   return GV;
827 }
828 
829 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
830                                         bool IsOne,
831                                         Constant *UniqueMemberAddr) {
832   for (auto &&Call : CSInfo.CallSites) {
833     IRBuilder<> B(Call.CS.getInstruction());
834     Value *Cmp = B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
835                               Call.VTable, UniqueMemberAddr);
836     Cmp = B.CreateZExt(Cmp, Call.CS->getType());
837     Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, Cmp);
838   }
839   CSInfo.markDevirt();
840 }
841 
842 bool DevirtModule::tryUniqueRetValOpt(
843     unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot,
844     CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res,
845     VTableSlot Slot, ArrayRef<uint64_t> Args) {
846   // IsOne controls whether we look for a 0 or a 1.
847   auto tryUniqueRetValOptFor = [&](bool IsOne) {
848     const TypeMemberInfo *UniqueMember = nullptr;
849     for (const VirtualCallTarget &Target : TargetsForSlot) {
850       if (Target.RetVal == (IsOne ? 1 : 0)) {
851         if (UniqueMember)
852           return false;
853         UniqueMember = Target.TM;
854       }
855     }
856 
857     // We should have found a unique member or bailed out by now. We already
858     // checked for a uniform return value in tryUniformRetValOpt.
859     assert(UniqueMember);
860 
861     Constant *UniqueMemberAddr =
862         ConstantExpr::getBitCast(UniqueMember->Bits->GV, Int8PtrTy);
863     UniqueMemberAddr = ConstantExpr::getGetElementPtr(
864         Int8Ty, UniqueMemberAddr,
865         ConstantInt::get(Int64Ty, UniqueMember->Offset));
866 
867     if (CSInfo.isExported()) {
868       Res->TheKind = WholeProgramDevirtResolution::ByArg::UniqueRetVal;
869       Res->Info = IsOne;
870 
871       exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr);
872     }
873 
874     // Replace each call with the comparison.
875     applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne,
876                          UniqueMemberAddr);
877 
878     // Update devirtualization statistics for targets.
879     if (RemarksEnabled)
880       for (auto &&Target : TargetsForSlot)
881         Target.WasDevirt = true;
882 
883     return true;
884   };
885 
886   if (BitWidth == 1) {
887     if (tryUniqueRetValOptFor(true))
888       return true;
889     if (tryUniqueRetValOptFor(false))
890       return true;
891   }
892   return false;
893 }
894 
895 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName,
896                                          Constant *Byte, Constant *Bit) {
897   for (auto Call : CSInfo.CallSites) {
898     auto *RetType = cast<IntegerType>(Call.CS.getType());
899     IRBuilder<> B(Call.CS.getInstruction());
900     Value *Addr = B.CreateGEP(Int8Ty, Call.VTable, Byte);
901     if (RetType->getBitWidth() == 1) {
902       Value *Bits = B.CreateLoad(Addr);
903       Value *BitsAndBit = B.CreateAnd(Bits, Bit);
904       auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0));
905       Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled,
906                            IsBitSet);
907     } else {
908       Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo());
909       Value *Val = B.CreateLoad(RetType, ValAddr);
910       Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled, Val);
911     }
912   }
913   CSInfo.markDevirt();
914 }
915 
916 bool DevirtModule::tryVirtualConstProp(
917     MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo,
918     WholeProgramDevirtResolution *Res, VTableSlot Slot) {
919   // This only works if the function returns an integer.
920   auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType());
921   if (!RetType)
922     return false;
923   unsigned BitWidth = RetType->getBitWidth();
924   if (BitWidth > 64)
925     return false;
926 
927   // Make sure that each function is defined, does not access memory, takes at
928   // least one argument, does not use its first argument (which we assume is
929   // 'this'), and has the same return type.
930   //
931   // Note that we test whether this copy of the function is readnone, rather
932   // than testing function attributes, which must hold for any copy of the
933   // function, even a less optimized version substituted at link time. This is
934   // sound because the virtual constant propagation optimizations effectively
935   // inline all implementations of the virtual function into each call site,
936   // rather than using function attributes to perform local optimization.
937   for (VirtualCallTarget &Target : TargetsForSlot) {
938     if (Target.Fn->isDeclaration() ||
939         computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) !=
940             MAK_ReadNone ||
941         Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() ||
942         Target.Fn->getReturnType() != RetType)
943       return false;
944   }
945 
946   for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) {
947     if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first))
948       continue;
949 
950     WholeProgramDevirtResolution::ByArg *ResByArg = nullptr;
951     if (Res)
952       ResByArg = &Res->ResByArg[CSByConstantArg.first];
953 
954     if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg))
955       continue;
956 
957     if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second,
958                            ResByArg, Slot, CSByConstantArg.first))
959       continue;
960 
961     // Find an allocation offset in bits in all vtables associated with the
962     // type.
963     uint64_t AllocBefore =
964         findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth);
965     uint64_t AllocAfter =
966         findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth);
967 
968     // Calculate the total amount of padding needed to store a value at both
969     // ends of the object.
970     uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0;
971     for (auto &&Target : TargetsForSlot) {
972       TotalPaddingBefore += std::max<int64_t>(
973           (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0);
974       TotalPaddingAfter += std::max<int64_t>(
975           (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0);
976     }
977 
978     // If the amount of padding is too large, give up.
979     // FIXME: do something smarter here.
980     if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128)
981       continue;
982 
983     // Calculate the offset to the value as a (possibly negative) byte offset
984     // and (if applicable) a bit offset, and store the values in the targets.
985     int64_t OffsetByte;
986     uint64_t OffsetBit;
987     if (TotalPaddingBefore <= TotalPaddingAfter)
988       setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte,
989                             OffsetBit);
990     else
991       setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte,
992                            OffsetBit);
993 
994     if (RemarksEnabled)
995       for (auto &&Target : TargetsForSlot)
996         Target.WasDevirt = true;
997 
998     Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte);
999     Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit);
1000 
1001     if (CSByConstantArg.second.isExported()) {
1002       ResByArg->TheKind = WholeProgramDevirtResolution::ByArg::VirtualConstProp;
1003       exportGlobal(Slot, CSByConstantArg.first, "byte",
1004                    ConstantExpr::getIntToPtr(ByteConst, Int8PtrTy));
1005       exportGlobal(Slot, CSByConstantArg.first, "bit",
1006                    ConstantExpr::getIntToPtr(BitConst, Int8PtrTy));
1007     }
1008 
1009     // Rewrite each call to a load from OffsetByte/OffsetBit.
1010     applyVirtualConstProp(CSByConstantArg.second,
1011                           TargetsForSlot[0].Fn->getName(), ByteConst, BitConst);
1012   }
1013   return true;
1014 }
1015 
1016 void DevirtModule::rebuildGlobal(VTableBits &B) {
1017   if (B.Before.Bytes.empty() && B.After.Bytes.empty())
1018     return;
1019 
1020   // Align each byte array to pointer width.
1021   unsigned PointerSize = M.getDataLayout().getPointerSize();
1022   B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), PointerSize));
1023   B.After.Bytes.resize(alignTo(B.After.Bytes.size(), PointerSize));
1024 
1025   // Before was stored in reverse order; flip it now.
1026   for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I)
1027     std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]);
1028 
1029   // Build an anonymous global containing the before bytes, followed by the
1030   // original initializer, followed by the after bytes.
1031   auto NewInit = ConstantStruct::getAnon(
1032       {ConstantDataArray::get(M.getContext(), B.Before.Bytes),
1033        B.GV->getInitializer(),
1034        ConstantDataArray::get(M.getContext(), B.After.Bytes)});
1035   auto NewGV =
1036       new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(),
1037                          GlobalVariable::PrivateLinkage, NewInit, "", B.GV);
1038   NewGV->setSection(B.GV->getSection());
1039   NewGV->setComdat(B.GV->getComdat());
1040 
1041   // Copy the original vtable's metadata to the anonymous global, adjusting
1042   // offsets as required.
1043   NewGV->copyMetadata(B.GV, B.Before.Bytes.size());
1044 
1045   // Build an alias named after the original global, pointing at the second
1046   // element (the original initializer).
1047   auto Alias = GlobalAlias::create(
1048       B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "",
1049       ConstantExpr::getGetElementPtr(
1050           NewInit->getType(), NewGV,
1051           ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0),
1052                                ConstantInt::get(Int32Ty, 1)}),
1053       &M);
1054   Alias->setVisibility(B.GV->getVisibility());
1055   Alias->takeName(B.GV);
1056 
1057   B.GV->replaceAllUsesWith(Alias);
1058   B.GV->eraseFromParent();
1059 }
1060 
1061 bool DevirtModule::areRemarksEnabled() {
1062   const auto &FL = M.getFunctionList();
1063   if (FL.empty())
1064     return false;
1065   const Function &Fn = FL.front();
1066 
1067   const auto &BBL = Fn.getBasicBlockList();
1068   if (BBL.empty())
1069     return false;
1070   auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front());
1071   return DI.isEnabled();
1072 }
1073 
1074 void DevirtModule::scanTypeTestUsers(Function *TypeTestFunc,
1075                                      Function *AssumeFunc) {
1076   // Find all virtual calls via a virtual table pointer %p under an assumption
1077   // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p
1078   // points to a member of the type identifier %md. Group calls by (type ID,
1079   // offset) pair (effectively the identity of the virtual function) and store
1080   // to CallSlots.
1081   DenseSet<Value *> SeenPtrs;
1082   for (auto I = TypeTestFunc->use_begin(), E = TypeTestFunc->use_end();
1083        I != E;) {
1084     auto CI = dyn_cast<CallInst>(I->getUser());
1085     ++I;
1086     if (!CI)
1087       continue;
1088 
1089     // Search for virtual calls based on %p and add them to DevirtCalls.
1090     SmallVector<DevirtCallSite, 1> DevirtCalls;
1091     SmallVector<CallInst *, 1> Assumes;
1092     findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI);
1093 
1094     // If we found any, add them to CallSlots. Only do this if we haven't seen
1095     // the vtable pointer before, as it may have been CSE'd with pointers from
1096     // other call sites, and we don't want to process call sites multiple times.
1097     if (!Assumes.empty()) {
1098       Metadata *TypeId =
1099           cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata();
1100       Value *Ptr = CI->getArgOperand(0)->stripPointerCasts();
1101       if (SeenPtrs.insert(Ptr).second) {
1102         for (DevirtCallSite Call : DevirtCalls) {
1103           CallSlots[{TypeId, Call.Offset}].addCallSite(CI->getArgOperand(0),
1104                                                        Call.CS, nullptr);
1105         }
1106       }
1107     }
1108 
1109     // We no longer need the assumes or the type test.
1110     for (auto Assume : Assumes)
1111       Assume->eraseFromParent();
1112     // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we
1113     // may use the vtable argument later.
1114     if (CI->use_empty())
1115       CI->eraseFromParent();
1116   }
1117 }
1118 
1119 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) {
1120   Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test);
1121 
1122   for (auto I = TypeCheckedLoadFunc->use_begin(),
1123             E = TypeCheckedLoadFunc->use_end();
1124        I != E;) {
1125     auto CI = dyn_cast<CallInst>(I->getUser());
1126     ++I;
1127     if (!CI)
1128       continue;
1129 
1130     Value *Ptr = CI->getArgOperand(0);
1131     Value *Offset = CI->getArgOperand(1);
1132     Value *TypeIdValue = CI->getArgOperand(2);
1133     Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata();
1134 
1135     SmallVector<DevirtCallSite, 1> DevirtCalls;
1136     SmallVector<Instruction *, 1> LoadedPtrs;
1137     SmallVector<Instruction *, 1> Preds;
1138     bool HasNonCallUses = false;
1139     findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds,
1140                                                HasNonCallUses, CI);
1141 
1142     // Start by generating "pessimistic" code that explicitly loads the function
1143     // pointer from the vtable and performs the type check. If possible, we will
1144     // eliminate the load and the type check later.
1145 
1146     // If possible, only generate the load at the point where it is used.
1147     // This helps avoid unnecessary spills.
1148     IRBuilder<> LoadB(
1149         (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI);
1150     Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset);
1151     Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy));
1152     Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr);
1153 
1154     for (Instruction *LoadedPtr : LoadedPtrs) {
1155       LoadedPtr->replaceAllUsesWith(LoadedValue);
1156       LoadedPtr->eraseFromParent();
1157     }
1158 
1159     // Likewise for the type test.
1160     IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI);
1161     CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue});
1162 
1163     for (Instruction *Pred : Preds) {
1164       Pred->replaceAllUsesWith(TypeTestCall);
1165       Pred->eraseFromParent();
1166     }
1167 
1168     // We have already erased any extractvalue instructions that refer to the
1169     // intrinsic call, but the intrinsic may have other non-extractvalue uses
1170     // (although this is unlikely). In that case, explicitly build a pair and
1171     // RAUW it.
1172     if (!CI->use_empty()) {
1173       Value *Pair = UndefValue::get(CI->getType());
1174       IRBuilder<> B(CI);
1175       Pair = B.CreateInsertValue(Pair, LoadedValue, {0});
1176       Pair = B.CreateInsertValue(Pair, TypeTestCall, {1});
1177       CI->replaceAllUsesWith(Pair);
1178     }
1179 
1180     // The number of unsafe uses is initially the number of uses.
1181     auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall];
1182     NumUnsafeUses = DevirtCalls.size();
1183 
1184     // If the function pointer has a non-call user, we cannot eliminate the type
1185     // check, as one of those users may eventually call the pointer. Increment
1186     // the unsafe use count to make sure it cannot reach zero.
1187     if (HasNonCallUses)
1188       ++NumUnsafeUses;
1189     for (DevirtCallSite Call : DevirtCalls) {
1190       CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS,
1191                                                    &NumUnsafeUses);
1192     }
1193 
1194     CI->eraseFromParent();
1195   }
1196 }
1197 
1198 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) {
1199   const WholeProgramDevirtResolution &Res =
1200       Summary->getTypeIdSummary(cast<MDString>(Slot.TypeID)->getString())
1201           .WPDRes[Slot.ByteOffset];
1202 
1203   if (Res.TheKind == WholeProgramDevirtResolution::SingleImpl) {
1204     // The type of the function in the declaration is irrelevant because every
1205     // call site will cast it to the correct type.
1206     auto *SingleImpl = M.getOrInsertFunction(
1207         Res.SingleImplName, Type::getVoidTy(M.getContext()), nullptr);
1208 
1209     // This is the import phase so we should not be exporting anything.
1210     bool IsExported = false;
1211     applySingleImplDevirt(SlotInfo, SingleImpl, IsExported);
1212     assert(!IsExported);
1213   }
1214 
1215   for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) {
1216     auto I = Res.ResByArg.find(CSByConstantArg.first);
1217     if (I == Res.ResByArg.end())
1218       continue;
1219     auto &ResByArg = I->second;
1220     // FIXME: We should figure out what to do about the "function name" argument
1221     // to the apply* functions, as the function names are unavailable during the
1222     // importing phase. For now we just pass the empty string. This does not
1223     // impact correctness because the function names are just used for remarks.
1224     switch (ResByArg.TheKind) {
1225     case WholeProgramDevirtResolution::ByArg::UniformRetVal:
1226       applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info);
1227       break;
1228     case WholeProgramDevirtResolution::ByArg::UniqueRetVal: {
1229       Constant *UniqueMemberAddr =
1230           importGlobal(Slot, CSByConstantArg.first, "unique_member");
1231       applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info,
1232                            UniqueMemberAddr);
1233       break;
1234     }
1235     case WholeProgramDevirtResolution::ByArg::VirtualConstProp: {
1236       Constant *Byte = importGlobal(Slot, CSByConstantArg.first, "byte", 32);
1237       Byte = ConstantExpr::getPtrToInt(Byte, Int32Ty);
1238       Constant *Bit = importGlobal(Slot, CSByConstantArg.first, "bit", 8);
1239       Bit = ConstantExpr::getPtrToInt(Bit, Int8Ty);
1240       applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit);
1241     }
1242     default:
1243       break;
1244     }
1245   }
1246 }
1247 
1248 void DevirtModule::removeRedundantTypeTests() {
1249   auto True = ConstantInt::getTrue(M.getContext());
1250   for (auto &&U : NumUnsafeUsesForTypeTest) {
1251     if (U.second == 0) {
1252       U.first->replaceAllUsesWith(True);
1253       U.first->eraseFromParent();
1254     }
1255   }
1256 }
1257 
1258 bool DevirtModule::run() {
1259   Function *TypeTestFunc =
1260       M.getFunction(Intrinsic::getName(Intrinsic::type_test));
1261   Function *TypeCheckedLoadFunc =
1262       M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load));
1263   Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume));
1264 
1265   // Normally if there are no users of the devirtualization intrinsics in the
1266   // module, this pass has nothing to do. But if we are exporting, we also need
1267   // to handle any users that appear only in the function summaries.
1268   if (Action != PassSummaryAction::Export &&
1269       (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc ||
1270        AssumeFunc->use_empty()) &&
1271       (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty()))
1272     return false;
1273 
1274   if (TypeTestFunc && AssumeFunc)
1275     scanTypeTestUsers(TypeTestFunc, AssumeFunc);
1276 
1277   if (TypeCheckedLoadFunc)
1278     scanTypeCheckedLoadUsers(TypeCheckedLoadFunc);
1279 
1280   if (Action == PassSummaryAction::Import) {
1281     for (auto &S : CallSlots)
1282       importResolution(S.first, S.second);
1283 
1284     removeRedundantTypeTests();
1285 
1286     // The rest of the code is only necessary when exporting or during regular
1287     // LTO, so we are done.
1288     return true;
1289   }
1290 
1291   // Rebuild type metadata into a map for easy lookup.
1292   std::vector<VTableBits> Bits;
1293   DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap;
1294   buildTypeIdentifierMap(Bits, TypeIdMap);
1295   if (TypeIdMap.empty())
1296     return true;
1297 
1298   // Collect information from summary about which calls to try to devirtualize.
1299   if (Action == PassSummaryAction::Export) {
1300     DenseMap<GlobalValue::GUID, TinyPtrVector<Metadata *>> MetadataByGUID;
1301     for (auto &P : TypeIdMap) {
1302       if (auto *TypeId = dyn_cast<MDString>(P.first))
1303         MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back(
1304             TypeId);
1305     }
1306 
1307     for (auto &P : *Summary) {
1308       for (auto &S : P.second) {
1309         auto *FS = dyn_cast<FunctionSummary>(S.get());
1310         if (!FS)
1311           continue;
1312         // FIXME: Only add live functions.
1313         for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) {
1314           for (Metadata *MD : MetadataByGUID[VF.GUID]) {
1315             CallSlots[{MD, VF.Offset}].CSInfo.SummaryHasTypeTestAssumeUsers =
1316                 true;
1317           }
1318         }
1319         for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) {
1320           for (Metadata *MD : MetadataByGUID[VF.GUID]) {
1321             CallSlots[{MD, VF.Offset}]
1322                 .CSInfo.SummaryTypeCheckedLoadUsers.push_back(FS);
1323           }
1324         }
1325         for (const FunctionSummary::ConstVCall &VC :
1326              FS->type_test_assume_const_vcalls()) {
1327           for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) {
1328             CallSlots[{MD, VC.VFunc.Offset}]
1329                 .ConstCSInfo[VC.Args]
1330                 .SummaryHasTypeTestAssumeUsers = true;
1331           }
1332         }
1333         for (const FunctionSummary::ConstVCall &VC :
1334              FS->type_checked_load_const_vcalls()) {
1335           for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) {
1336             CallSlots[{MD, VC.VFunc.Offset}]
1337                 .ConstCSInfo[VC.Args]
1338                 .SummaryTypeCheckedLoadUsers.push_back(FS);
1339           }
1340         }
1341       }
1342     }
1343   }
1344 
1345   // For each (type, offset) pair:
1346   bool DidVirtualConstProp = false;
1347   std::map<std::string, Function*> DevirtTargets;
1348   for (auto &S : CallSlots) {
1349     // Search each of the members of the type identifier for the virtual
1350     // function implementation at offset S.first.ByteOffset, and add to
1351     // TargetsForSlot.
1352     std::vector<VirtualCallTarget> TargetsForSlot;
1353     if (tryFindVirtualCallTargets(TargetsForSlot, TypeIdMap[S.first.TypeID],
1354                                   S.first.ByteOffset)) {
1355       WholeProgramDevirtResolution *Res = nullptr;
1356       if (Action == PassSummaryAction::Export && isa<MDString>(S.first.TypeID))
1357         Res =
1358             &Summary
1359                  ->getTypeIdSummary(cast<MDString>(S.first.TypeID)->getString())
1360                  .WPDRes[S.first.ByteOffset];
1361 
1362       if (!trySingleImplDevirt(TargetsForSlot, S.second, Res) &&
1363           tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first))
1364         DidVirtualConstProp = true;
1365 
1366       // Collect functions devirtualized at least for one call site for stats.
1367       if (RemarksEnabled)
1368         for (const auto &T : TargetsForSlot)
1369           if (T.WasDevirt)
1370             DevirtTargets[T.Fn->getName()] = T.Fn;
1371     }
1372 
1373     // CFI-specific: if we are exporting and any llvm.type.checked.load
1374     // intrinsics were *not* devirtualized, we need to add the resulting
1375     // llvm.type.test intrinsics to the function summaries so that the
1376     // LowerTypeTests pass will export them.
1377     if (Action == PassSummaryAction::Export && isa<MDString>(S.first.TypeID)) {
1378       auto GUID =
1379           GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString());
1380       for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers)
1381         FS->addTypeTest(GUID);
1382       for (auto &CCS : S.second.ConstCSInfo)
1383         for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers)
1384           FS->addTypeTest(GUID);
1385     }
1386   }
1387 
1388   if (RemarksEnabled) {
1389     // Generate remarks for each devirtualized function.
1390     for (const auto &DT : DevirtTargets) {
1391       Function *F = DT.second;
1392       DISubprogram *SP = F->getSubprogram();
1393       emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, SP,
1394                              Twine("devirtualized ") + F->getName());
1395     }
1396   }
1397 
1398   removeRedundantTypeTests();
1399 
1400   // Rebuild each global we touched as part of virtual constant propagation to
1401   // include the before and after bytes.
1402   if (DidVirtualConstProp)
1403     for (VTableBits &B : Bits)
1404       rebuildGlobal(B);
1405 
1406   return true;
1407 }
1408