1 //===- WholeProgramDevirt.cpp - Whole program virtual call optimization ---===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass implements whole program optimization of virtual calls in cases
10 // where we know (via !type metadata) that the list of callees is fixed. This
11 // includes the following:
12 // - Single implementation devirtualization: if a virtual call has a single
13 //   possible callee, replace all calls with a direct call to that callee.
14 // - Virtual constant propagation: if the virtual function's return type is an
15 //   integer <=64 bits and all possible callees are readnone, for each class and
16 //   each list of constant arguments: evaluate the function, store the return
17 //   value alongside the virtual table, and rewrite each virtual call as a load
18 //   from the virtual table.
19 // - Uniform return value optimization: if the conditions for virtual constant
20 //   propagation hold and each function returns the same constant value, replace
21 //   each virtual call with that constant.
22 // - Unique return value optimization for i1 return values: if the conditions
23 //   for virtual constant propagation hold and a single vtable's function
24 //   returns 0, or a single vtable's function returns 1, replace each virtual
25 //   call with a comparison of the vptr against that vtable's address.
26 //
27 // This pass is intended to be used during the regular and thin LTO pipelines:
28 //
29 // During regular LTO, the pass determines the best optimization for each
30 // virtual call and applies the resolutions directly to virtual calls that are
31 // eligible for virtual call optimization (i.e. calls that use either of the
32 // llvm.assume(llvm.type.test) or llvm.type.checked.load intrinsics).
33 //
34 // During hybrid Regular/ThinLTO, the pass operates in two phases:
35 // - Export phase: this is run during the thin link over a single merged module
36 //   that contains all vtables with !type metadata that participate in the link.
37 //   The pass computes a resolution for each virtual call and stores it in the
38 //   type identifier summary.
39 // - Import phase: this is run during the thin backends over the individual
40 //   modules. The pass applies the resolutions previously computed during the
41 //   import phase to each eligible virtual call.
42 //
43 // During ThinLTO, the pass operates in two phases:
44 // - Export phase: this is run during the thin link over the index which
45 //   contains a summary of all vtables with !type metadata that participate in
46 //   the link. It computes a resolution for each virtual call and stores it in
47 //   the type identifier summary. Only single implementation devirtualization
48 //   is supported.
49 // - Import phase: (same as with hybrid case above).
50 //
51 //===----------------------------------------------------------------------===//
52 
53 #include "llvm/Transforms/IPO/WholeProgramDevirt.h"
54 #include "llvm/ADT/ArrayRef.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/DenseMapInfo.h"
57 #include "llvm/ADT/DenseSet.h"
58 #include "llvm/ADT/MapVector.h"
59 #include "llvm/ADT/SmallVector.h"
60 #include "llvm/ADT/iterator_range.h"
61 #include "llvm/Analysis/AliasAnalysis.h"
62 #include "llvm/Analysis/BasicAliasAnalysis.h"
63 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
64 #include "llvm/Analysis/TypeMetadataUtils.h"
65 #include "llvm/IR/CallSite.h"
66 #include "llvm/IR/Constants.h"
67 #include "llvm/IR/DataLayout.h"
68 #include "llvm/IR/DebugLoc.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/Dominators.h"
71 #include "llvm/IR/Function.h"
72 #include "llvm/IR/GlobalAlias.h"
73 #include "llvm/IR/GlobalVariable.h"
74 #include "llvm/IR/IRBuilder.h"
75 #include "llvm/IR/InstrTypes.h"
76 #include "llvm/IR/Instruction.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/Intrinsics.h"
79 #include "llvm/IR/LLVMContext.h"
80 #include "llvm/IR/Metadata.h"
81 #include "llvm/IR/Module.h"
82 #include "llvm/IR/ModuleSummaryIndexYAML.h"
83 #include "llvm/InitializePasses.h"
84 #include "llvm/Pass.h"
85 #include "llvm/PassRegistry.h"
86 #include "llvm/PassSupport.h"
87 #include "llvm/Support/Casting.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/Error.h"
90 #include "llvm/Support/FileSystem.h"
91 #include "llvm/Support/MathExtras.h"
92 #include "llvm/Transforms/IPO.h"
93 #include "llvm/Transforms/IPO/FunctionAttrs.h"
94 #include "llvm/Transforms/Utils/Evaluator.h"
95 #include <algorithm>
96 #include <cstddef>
97 #include <map>
98 #include <set>
99 #include <string>
100 
101 using namespace llvm;
102 using namespace wholeprogramdevirt;
103 
104 #define DEBUG_TYPE "wholeprogramdevirt"
105 
106 static cl::opt<PassSummaryAction> ClSummaryAction(
107     "wholeprogramdevirt-summary-action",
108     cl::desc("What to do with the summary when running this pass"),
109     cl::values(clEnumValN(PassSummaryAction::None, "none", "Do nothing"),
110                clEnumValN(PassSummaryAction::Import, "import",
111                           "Import typeid resolutions from summary and globals"),
112                clEnumValN(PassSummaryAction::Export, "export",
113                           "Export typeid resolutions to summary and globals")),
114     cl::Hidden);
115 
116 static cl::opt<std::string> ClReadSummary(
117     "wholeprogramdevirt-read-summary",
118     cl::desc("Read summary from given YAML file before running pass"),
119     cl::Hidden);
120 
121 static cl::opt<std::string> ClWriteSummary(
122     "wholeprogramdevirt-write-summary",
123     cl::desc("Write summary to given YAML file after running pass"),
124     cl::Hidden);
125 
126 static cl::opt<unsigned>
127     ClThreshold("wholeprogramdevirt-branch-funnel-threshold", cl::Hidden,
128                 cl::init(10), cl::ZeroOrMore,
129                 cl::desc("Maximum number of call targets per "
130                          "call site to enable branch funnels"));
131 
132 static cl::opt<bool>
133     PrintSummaryDevirt("wholeprogramdevirt-print-index-based", cl::Hidden,
134                        cl::init(false), cl::ZeroOrMore,
135                        cl::desc("Print index-based devirtualization messages"));
136 
137 // Find the minimum offset that we may store a value of size Size bits at. If
138 // IsAfter is set, look for an offset before the object, otherwise look for an
139 // offset after the object.
140 uint64_t
141 wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets,
142                                      bool IsAfter, uint64_t Size) {
143   // Find a minimum offset taking into account only vtable sizes.
144   uint64_t MinByte = 0;
145   for (const VirtualCallTarget &Target : Targets) {
146     if (IsAfter)
147       MinByte = std::max(MinByte, Target.minAfterBytes());
148     else
149       MinByte = std::max(MinByte, Target.minBeforeBytes());
150   }
151 
152   // Build a vector of arrays of bytes covering, for each target, a slice of the
153   // used region (see AccumBitVector::BytesUsed in
154   // llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively,
155   // this aligns the used regions to start at MinByte.
156   //
157   // In this example, A, B and C are vtables, # is a byte already allocated for
158   // a virtual function pointer, AAAA... (etc.) are the used regions for the
159   // vtables and Offset(X) is the value computed for the Offset variable below
160   // for X.
161   //
162   //                    Offset(A)
163   //                    |       |
164   //                            |MinByte
165   // A: ################AAAAAAAA|AAAAAAAA
166   // B: ########BBBBBBBBBBBBBBBB|BBBB
167   // C: ########################|CCCCCCCCCCCCCCCC
168   //            |   Offset(B)   |
169   //
170   // This code produces the slices of A, B and C that appear after the divider
171   // at MinByte.
172   std::vector<ArrayRef<uint8_t>> Used;
173   for (const VirtualCallTarget &Target : Targets) {
174     ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed
175                                        : Target.TM->Bits->Before.BytesUsed;
176     uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes()
177                               : MinByte - Target.minBeforeBytes();
178 
179     // Disregard used regions that are smaller than Offset. These are
180     // effectively all-free regions that do not need to be checked.
181     if (VTUsed.size() > Offset)
182       Used.push_back(VTUsed.slice(Offset));
183   }
184 
185   if (Size == 1) {
186     // Find a free bit in each member of Used.
187     for (unsigned I = 0;; ++I) {
188       uint8_t BitsUsed = 0;
189       for (auto &&B : Used)
190         if (I < B.size())
191           BitsUsed |= B[I];
192       if (BitsUsed != 0xff)
193         return (MinByte + I) * 8 +
194                countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined);
195     }
196   } else {
197     // Find a free (Size/8) byte region in each member of Used.
198     // FIXME: see if alignment helps.
199     for (unsigned I = 0;; ++I) {
200       for (auto &&B : Used) {
201         unsigned Byte = 0;
202         while ((I + Byte) < B.size() && Byte < (Size / 8)) {
203           if (B[I + Byte])
204             goto NextI;
205           ++Byte;
206         }
207       }
208       return (MinByte + I) * 8;
209     NextI:;
210     }
211   }
212 }
213 
214 void wholeprogramdevirt::setBeforeReturnValues(
215     MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore,
216     unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
217   if (BitWidth == 1)
218     OffsetByte = -(AllocBefore / 8 + 1);
219   else
220     OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8);
221   OffsetBit = AllocBefore % 8;
222 
223   for (VirtualCallTarget &Target : Targets) {
224     if (BitWidth == 1)
225       Target.setBeforeBit(AllocBefore);
226     else
227       Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8);
228   }
229 }
230 
231 void wholeprogramdevirt::setAfterReturnValues(
232     MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter,
233     unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
234   if (BitWidth == 1)
235     OffsetByte = AllocAfter / 8;
236   else
237     OffsetByte = (AllocAfter + 7) / 8;
238   OffsetBit = AllocAfter % 8;
239 
240   for (VirtualCallTarget &Target : Targets) {
241     if (BitWidth == 1)
242       Target.setAfterBit(AllocAfter);
243     else
244       Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8);
245   }
246 }
247 
248 VirtualCallTarget::VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM)
249     : Fn(Fn), TM(TM),
250       IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()), WasDevirt(false) {}
251 
252 namespace {
253 
254 // A slot in a set of virtual tables. The TypeID identifies the set of virtual
255 // tables, and the ByteOffset is the offset in bytes from the address point to
256 // the virtual function pointer.
257 struct VTableSlot {
258   Metadata *TypeID;
259   uint64_t ByteOffset;
260 };
261 
262 } // end anonymous namespace
263 
264 namespace llvm {
265 
266 template <> struct DenseMapInfo<VTableSlot> {
267   static VTableSlot getEmptyKey() {
268     return {DenseMapInfo<Metadata *>::getEmptyKey(),
269             DenseMapInfo<uint64_t>::getEmptyKey()};
270   }
271   static VTableSlot getTombstoneKey() {
272     return {DenseMapInfo<Metadata *>::getTombstoneKey(),
273             DenseMapInfo<uint64_t>::getTombstoneKey()};
274   }
275   static unsigned getHashValue(const VTableSlot &I) {
276     return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^
277            DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset);
278   }
279   static bool isEqual(const VTableSlot &LHS,
280                       const VTableSlot &RHS) {
281     return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset;
282   }
283 };
284 
285 template <> struct DenseMapInfo<VTableSlotSummary> {
286   static VTableSlotSummary getEmptyKey() {
287     return {DenseMapInfo<StringRef>::getEmptyKey(),
288             DenseMapInfo<uint64_t>::getEmptyKey()};
289   }
290   static VTableSlotSummary getTombstoneKey() {
291     return {DenseMapInfo<StringRef>::getTombstoneKey(),
292             DenseMapInfo<uint64_t>::getTombstoneKey()};
293   }
294   static unsigned getHashValue(const VTableSlotSummary &I) {
295     return DenseMapInfo<StringRef>::getHashValue(I.TypeID) ^
296            DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset);
297   }
298   static bool isEqual(const VTableSlotSummary &LHS,
299                       const VTableSlotSummary &RHS) {
300     return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset;
301   }
302 };
303 
304 } // end namespace llvm
305 
306 namespace {
307 
308 // A virtual call site. VTable is the loaded virtual table pointer, and CS is
309 // the indirect virtual call.
310 struct VirtualCallSite {
311   Value *VTable;
312   CallSite CS;
313 
314   // If non-null, this field points to the associated unsafe use count stored in
315   // the DevirtModule::NumUnsafeUsesForTypeTest map below. See the description
316   // of that field for details.
317   unsigned *NumUnsafeUses;
318 
319   void
320   emitRemark(const StringRef OptName, const StringRef TargetName,
321              function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter) {
322     Function *F = CS.getCaller();
323     DebugLoc DLoc = CS->getDebugLoc();
324     BasicBlock *Block = CS.getParent();
325 
326     using namespace ore;
327     OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, OptName, DLoc, Block)
328                       << NV("Optimization", OptName)
329                       << ": devirtualized a call to "
330                       << NV("FunctionName", TargetName));
331   }
332 
333   void replaceAndErase(
334       const StringRef OptName, const StringRef TargetName, bool RemarksEnabled,
335       function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter,
336       Value *New) {
337     if (RemarksEnabled)
338       emitRemark(OptName, TargetName, OREGetter);
339     CS->replaceAllUsesWith(New);
340     if (auto II = dyn_cast<InvokeInst>(CS.getInstruction())) {
341       BranchInst::Create(II->getNormalDest(), CS.getInstruction());
342       II->getUnwindDest()->removePredecessor(II->getParent());
343     }
344     CS->eraseFromParent();
345     // This use is no longer unsafe.
346     if (NumUnsafeUses)
347       --*NumUnsafeUses;
348   }
349 };
350 
351 // Call site information collected for a specific VTableSlot and possibly a list
352 // of constant integer arguments. The grouping by arguments is handled by the
353 // VTableSlotInfo class.
354 struct CallSiteInfo {
355   /// The set of call sites for this slot. Used during regular LTO and the
356   /// import phase of ThinLTO (as well as the export phase of ThinLTO for any
357   /// call sites that appear in the merged module itself); in each of these
358   /// cases we are directly operating on the call sites at the IR level.
359   std::vector<VirtualCallSite> CallSites;
360 
361   /// Whether all call sites represented by this CallSiteInfo, including those
362   /// in summaries, have been devirtualized. This starts off as true because a
363   /// default constructed CallSiteInfo represents no call sites.
364   bool AllCallSitesDevirted = true;
365 
366   // These fields are used during the export phase of ThinLTO and reflect
367   // information collected from function summaries.
368 
369   /// Whether any function summary contains an llvm.assume(llvm.type.test) for
370   /// this slot.
371   bool SummaryHasTypeTestAssumeUsers = false;
372 
373   /// CFI-specific: a vector containing the list of function summaries that use
374   /// the llvm.type.checked.load intrinsic and therefore will require
375   /// resolutions for llvm.type.test in order to implement CFI checks if
376   /// devirtualization was unsuccessful. If devirtualization was successful, the
377   /// pass will clear this vector by calling markDevirt(). If at the end of the
378   /// pass the vector is non-empty, we will need to add a use of llvm.type.test
379   /// to each of the function summaries in the vector.
380   std::vector<FunctionSummary *> SummaryTypeCheckedLoadUsers;
381   std::vector<FunctionSummary *> SummaryTypeTestAssumeUsers;
382 
383   bool isExported() const {
384     return SummaryHasTypeTestAssumeUsers ||
385            !SummaryTypeCheckedLoadUsers.empty();
386   }
387 
388   void addSummaryTypeCheckedLoadUser(FunctionSummary *FS) {
389     SummaryTypeCheckedLoadUsers.push_back(FS);
390     AllCallSitesDevirted = false;
391   }
392 
393   void addSummaryTypeTestAssumeUser(FunctionSummary *FS) {
394     SummaryTypeTestAssumeUsers.push_back(FS);
395     SummaryHasTypeTestAssumeUsers = true;
396     AllCallSitesDevirted = false;
397   }
398 
399   void markDevirt() {
400     AllCallSitesDevirted = true;
401 
402     // As explained in the comment for SummaryTypeCheckedLoadUsers.
403     SummaryTypeCheckedLoadUsers.clear();
404   }
405 };
406 
407 // Call site information collected for a specific VTableSlot.
408 struct VTableSlotInfo {
409   // The set of call sites which do not have all constant integer arguments
410   // (excluding "this").
411   CallSiteInfo CSInfo;
412 
413   // The set of call sites with all constant integer arguments (excluding
414   // "this"), grouped by argument list.
415   std::map<std::vector<uint64_t>, CallSiteInfo> ConstCSInfo;
416 
417   void addCallSite(Value *VTable, CallSite CS, unsigned *NumUnsafeUses);
418 
419 private:
420   CallSiteInfo &findCallSiteInfo(CallSite CS);
421 };
422 
423 CallSiteInfo &VTableSlotInfo::findCallSiteInfo(CallSite CS) {
424   std::vector<uint64_t> Args;
425   auto *CI = dyn_cast<IntegerType>(CS.getType());
426   if (!CI || CI->getBitWidth() > 64 || CS.arg_empty())
427     return CSInfo;
428   for (auto &&Arg : make_range(CS.arg_begin() + 1, CS.arg_end())) {
429     auto *CI = dyn_cast<ConstantInt>(Arg);
430     if (!CI || CI->getBitWidth() > 64)
431       return CSInfo;
432     Args.push_back(CI->getZExtValue());
433   }
434   return ConstCSInfo[Args];
435 }
436 
437 void VTableSlotInfo::addCallSite(Value *VTable, CallSite CS,
438                                  unsigned *NumUnsafeUses) {
439   auto &CSI = findCallSiteInfo(CS);
440   CSI.AllCallSitesDevirted = false;
441   CSI.CallSites.push_back({VTable, CS, NumUnsafeUses});
442 }
443 
444 struct DevirtModule {
445   Module &M;
446   function_ref<AAResults &(Function &)> AARGetter;
447   function_ref<DominatorTree &(Function &)> LookupDomTree;
448 
449   ModuleSummaryIndex *ExportSummary;
450   const ModuleSummaryIndex *ImportSummary;
451 
452   IntegerType *Int8Ty;
453   PointerType *Int8PtrTy;
454   IntegerType *Int32Ty;
455   IntegerType *Int64Ty;
456   IntegerType *IntPtrTy;
457 
458   bool RemarksEnabled;
459   function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter;
460 
461   MapVector<VTableSlot, VTableSlotInfo> CallSlots;
462 
463   // This map keeps track of the number of "unsafe" uses of a loaded function
464   // pointer. The key is the associated llvm.type.test intrinsic call generated
465   // by this pass. An unsafe use is one that calls the loaded function pointer
466   // directly. Every time we eliminate an unsafe use (for example, by
467   // devirtualizing it or by applying virtual constant propagation), we
468   // decrement the value stored in this map. If a value reaches zero, we can
469   // eliminate the type check by RAUWing the associated llvm.type.test call with
470   // true.
471   std::map<CallInst *, unsigned> NumUnsafeUsesForTypeTest;
472 
473   DevirtModule(Module &M, function_ref<AAResults &(Function &)> AARGetter,
474                function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter,
475                function_ref<DominatorTree &(Function &)> LookupDomTree,
476                ModuleSummaryIndex *ExportSummary,
477                const ModuleSummaryIndex *ImportSummary)
478       : M(M), AARGetter(AARGetter), LookupDomTree(LookupDomTree),
479         ExportSummary(ExportSummary), ImportSummary(ImportSummary),
480         Int8Ty(Type::getInt8Ty(M.getContext())),
481         Int8PtrTy(Type::getInt8PtrTy(M.getContext())),
482         Int32Ty(Type::getInt32Ty(M.getContext())),
483         Int64Ty(Type::getInt64Ty(M.getContext())),
484         IntPtrTy(M.getDataLayout().getIntPtrType(M.getContext(), 0)),
485         RemarksEnabled(areRemarksEnabled()), OREGetter(OREGetter) {
486     assert(!(ExportSummary && ImportSummary));
487   }
488 
489   bool areRemarksEnabled();
490 
491   void scanTypeTestUsers(Function *TypeTestFunc);
492   void scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc);
493 
494   void buildTypeIdentifierMap(
495       std::vector<VTableBits> &Bits,
496       DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap);
497   bool
498   tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot,
499                             const std::set<TypeMemberInfo> &TypeMemberInfos,
500                             uint64_t ByteOffset);
501 
502   void applySingleImplDevirt(VTableSlotInfo &SlotInfo, Constant *TheFn,
503                              bool &IsExported);
504   bool trySingleImplDevirt(ModuleSummaryIndex *ExportSummary,
505                            MutableArrayRef<VirtualCallTarget> TargetsForSlot,
506                            VTableSlotInfo &SlotInfo,
507                            WholeProgramDevirtResolution *Res);
508 
509   void applyICallBranchFunnel(VTableSlotInfo &SlotInfo, Constant *JT,
510                               bool &IsExported);
511   void tryICallBranchFunnel(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
512                             VTableSlotInfo &SlotInfo,
513                             WholeProgramDevirtResolution *Res, VTableSlot Slot);
514 
515   bool tryEvaluateFunctionsWithArgs(
516       MutableArrayRef<VirtualCallTarget> TargetsForSlot,
517       ArrayRef<uint64_t> Args);
518 
519   void applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
520                              uint64_t TheRetVal);
521   bool tryUniformRetValOpt(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
522                            CallSiteInfo &CSInfo,
523                            WholeProgramDevirtResolution::ByArg *Res);
524 
525   // Returns the global symbol name that is used to export information about the
526   // given vtable slot and list of arguments.
527   std::string getGlobalName(VTableSlot Slot, ArrayRef<uint64_t> Args,
528                             StringRef Name);
529 
530   bool shouldExportConstantsAsAbsoluteSymbols();
531 
532   // This function is called during the export phase to create a symbol
533   // definition containing information about the given vtable slot and list of
534   // arguments.
535   void exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name,
536                     Constant *C);
537   void exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args, StringRef Name,
538                       uint32_t Const, uint32_t &Storage);
539 
540   // This function is called during the import phase to create a reference to
541   // the symbol definition created during the export phase.
542   Constant *importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
543                          StringRef Name);
544   Constant *importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
545                            StringRef Name, IntegerType *IntTy,
546                            uint32_t Storage);
547 
548   Constant *getMemberAddr(const TypeMemberInfo *M);
549 
550   void applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName, bool IsOne,
551                             Constant *UniqueMemberAddr);
552   bool tryUniqueRetValOpt(unsigned BitWidth,
553                           MutableArrayRef<VirtualCallTarget> TargetsForSlot,
554                           CallSiteInfo &CSInfo,
555                           WholeProgramDevirtResolution::ByArg *Res,
556                           VTableSlot Slot, ArrayRef<uint64_t> Args);
557 
558   void applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName,
559                              Constant *Byte, Constant *Bit);
560   bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
561                            VTableSlotInfo &SlotInfo,
562                            WholeProgramDevirtResolution *Res, VTableSlot Slot);
563 
564   void rebuildGlobal(VTableBits &B);
565 
566   // Apply the summary resolution for Slot to all virtual calls in SlotInfo.
567   void importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo);
568 
569   // If we were able to eliminate all unsafe uses for a type checked load,
570   // eliminate the associated type tests by replacing them with true.
571   void removeRedundantTypeTests();
572 
573   bool run();
574 
575   // Lower the module using the action and summary passed as command line
576   // arguments. For testing purposes only.
577   static bool
578   runForTesting(Module &M, function_ref<AAResults &(Function &)> AARGetter,
579                 function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter,
580                 function_ref<DominatorTree &(Function &)> LookupDomTree);
581 };
582 
583 struct DevirtIndex {
584   ModuleSummaryIndex &ExportSummary;
585   // The set in which to record GUIDs exported from their module by
586   // devirtualization, used by client to ensure they are not internalized.
587   std::set<GlobalValue::GUID> &ExportedGUIDs;
588   // A map in which to record the information necessary to locate the WPD
589   // resolution for local targets in case they are exported by cross module
590   // importing.
591   std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap;
592 
593   MapVector<VTableSlotSummary, VTableSlotInfo> CallSlots;
594 
595   DevirtIndex(
596       ModuleSummaryIndex &ExportSummary,
597       std::set<GlobalValue::GUID> &ExportedGUIDs,
598       std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap)
599       : ExportSummary(ExportSummary), ExportedGUIDs(ExportedGUIDs),
600         LocalWPDTargetsMap(LocalWPDTargetsMap) {}
601 
602   bool tryFindVirtualCallTargets(std::vector<ValueInfo> &TargetsForSlot,
603                                  const TypeIdCompatibleVtableInfo TIdInfo,
604                                  uint64_t ByteOffset);
605 
606   bool trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot,
607                            VTableSlotSummary &SlotSummary,
608                            VTableSlotInfo &SlotInfo,
609                            WholeProgramDevirtResolution *Res,
610                            std::set<ValueInfo> &DevirtTargets);
611 
612   void run();
613 };
614 
615 struct WholeProgramDevirt : public ModulePass {
616   static char ID;
617 
618   bool UseCommandLine = false;
619 
620   ModuleSummaryIndex *ExportSummary = nullptr;
621   const ModuleSummaryIndex *ImportSummary = nullptr;
622 
623   WholeProgramDevirt() : ModulePass(ID), UseCommandLine(true) {
624     initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry());
625   }
626 
627   WholeProgramDevirt(ModuleSummaryIndex *ExportSummary,
628                      const ModuleSummaryIndex *ImportSummary)
629       : ModulePass(ID), ExportSummary(ExportSummary),
630         ImportSummary(ImportSummary) {
631     initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry());
632   }
633 
634   bool runOnModule(Module &M) override {
635     if (skipModule(M))
636       return false;
637 
638     // In the new pass manager, we can request the optimization
639     // remark emitter pass on a per-function-basis, which the
640     // OREGetter will do for us.
641     // In the old pass manager, this is harder, so we just build
642     // an optimization remark emitter on the fly, when we need it.
643     std::unique_ptr<OptimizationRemarkEmitter> ORE;
644     auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & {
645       ORE = std::make_unique<OptimizationRemarkEmitter>(F);
646       return *ORE;
647     };
648 
649     auto LookupDomTree = [this](Function &F) -> DominatorTree & {
650       return this->getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
651     };
652 
653     if (UseCommandLine)
654       return DevirtModule::runForTesting(M, LegacyAARGetter(*this), OREGetter,
655                                          LookupDomTree);
656 
657     return DevirtModule(M, LegacyAARGetter(*this), OREGetter, LookupDomTree,
658                         ExportSummary, ImportSummary)
659         .run();
660   }
661 
662   void getAnalysisUsage(AnalysisUsage &AU) const override {
663     AU.addRequired<AssumptionCacheTracker>();
664     AU.addRequired<TargetLibraryInfoWrapperPass>();
665     AU.addRequired<DominatorTreeWrapperPass>();
666   }
667 };
668 
669 } // end anonymous namespace
670 
671 INITIALIZE_PASS_BEGIN(WholeProgramDevirt, "wholeprogramdevirt",
672                       "Whole program devirtualization", false, false)
673 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
674 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
675 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
676 INITIALIZE_PASS_END(WholeProgramDevirt, "wholeprogramdevirt",
677                     "Whole program devirtualization", false, false)
678 char WholeProgramDevirt::ID = 0;
679 
680 ModulePass *
681 llvm::createWholeProgramDevirtPass(ModuleSummaryIndex *ExportSummary,
682                                    const ModuleSummaryIndex *ImportSummary) {
683   return new WholeProgramDevirt(ExportSummary, ImportSummary);
684 }
685 
686 PreservedAnalyses WholeProgramDevirtPass::run(Module &M,
687                                               ModuleAnalysisManager &AM) {
688   auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
689   auto AARGetter = [&](Function &F) -> AAResults & {
690     return FAM.getResult<AAManager>(F);
691   };
692   auto OREGetter = [&](Function *F) -> OptimizationRemarkEmitter & {
693     return FAM.getResult<OptimizationRemarkEmitterAnalysis>(*F);
694   };
695   auto LookupDomTree = [&FAM](Function &F) -> DominatorTree & {
696     return FAM.getResult<DominatorTreeAnalysis>(F);
697   };
698   if (!DevirtModule(M, AARGetter, OREGetter, LookupDomTree, ExportSummary,
699                     ImportSummary)
700            .run())
701     return PreservedAnalyses::all();
702   return PreservedAnalyses::none();
703 }
704 
705 namespace llvm {
706 void runWholeProgramDevirtOnIndex(
707     ModuleSummaryIndex &Summary, std::set<GlobalValue::GUID> &ExportedGUIDs,
708     std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) {
709   DevirtIndex(Summary, ExportedGUIDs, LocalWPDTargetsMap).run();
710 }
711 
712 void updateIndexWPDForExports(
713     ModuleSummaryIndex &Summary,
714     function_ref<bool(StringRef, ValueInfo)> isExported,
715     std::map<ValueInfo, std::vector<VTableSlotSummary>> &LocalWPDTargetsMap) {
716   for (auto &T : LocalWPDTargetsMap) {
717     auto &VI = T.first;
718     // This was enforced earlier during trySingleImplDevirt.
719     assert(VI.getSummaryList().size() == 1 &&
720            "Devirt of local target has more than one copy");
721     auto &S = VI.getSummaryList()[0];
722     if (!isExported(S->modulePath(), VI))
723       continue;
724 
725     // It's been exported by a cross module import.
726     for (auto &SlotSummary : T.second) {
727       auto *TIdSum = Summary.getTypeIdSummary(SlotSummary.TypeID);
728       assert(TIdSum);
729       auto WPDRes = TIdSum->WPDRes.find(SlotSummary.ByteOffset);
730       assert(WPDRes != TIdSum->WPDRes.end());
731       WPDRes->second.SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal(
732           WPDRes->second.SingleImplName,
733           Summary.getModuleHash(S->modulePath()));
734     }
735   }
736 }
737 
738 } // end namespace llvm
739 
740 bool DevirtModule::runForTesting(
741     Module &M, function_ref<AAResults &(Function &)> AARGetter,
742     function_ref<OptimizationRemarkEmitter &(Function *)> OREGetter,
743     function_ref<DominatorTree &(Function &)> LookupDomTree) {
744   ModuleSummaryIndex Summary(/*HaveGVs=*/false);
745 
746   // Handle the command-line summary arguments. This code is for testing
747   // purposes only, so we handle errors directly.
748   if (!ClReadSummary.empty()) {
749     ExitOnError ExitOnErr("-wholeprogramdevirt-read-summary: " + ClReadSummary +
750                           ": ");
751     auto ReadSummaryFile =
752         ExitOnErr(errorOrToExpected(MemoryBuffer::getFile(ClReadSummary)));
753 
754     yaml::Input In(ReadSummaryFile->getBuffer());
755     In >> Summary;
756     ExitOnErr(errorCodeToError(In.error()));
757   }
758 
759   bool Changed =
760       DevirtModule(
761           M, AARGetter, OREGetter, LookupDomTree,
762           ClSummaryAction == PassSummaryAction::Export ? &Summary : nullptr,
763           ClSummaryAction == PassSummaryAction::Import ? &Summary : nullptr)
764           .run();
765 
766   if (!ClWriteSummary.empty()) {
767     ExitOnError ExitOnErr(
768         "-wholeprogramdevirt-write-summary: " + ClWriteSummary + ": ");
769     std::error_code EC;
770     raw_fd_ostream OS(ClWriteSummary, EC, sys::fs::OF_Text);
771     ExitOnErr(errorCodeToError(EC));
772 
773     yaml::Output Out(OS);
774     Out << Summary;
775   }
776 
777   return Changed;
778 }
779 
780 void DevirtModule::buildTypeIdentifierMap(
781     std::vector<VTableBits> &Bits,
782     DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) {
783   DenseMap<GlobalVariable *, VTableBits *> GVToBits;
784   Bits.reserve(M.getGlobalList().size());
785   SmallVector<MDNode *, 2> Types;
786   for (GlobalVariable &GV : M.globals()) {
787     Types.clear();
788     GV.getMetadata(LLVMContext::MD_type, Types);
789     if (GV.isDeclaration() || Types.empty())
790       continue;
791 
792     VTableBits *&BitsPtr = GVToBits[&GV];
793     if (!BitsPtr) {
794       Bits.emplace_back();
795       Bits.back().GV = &GV;
796       Bits.back().ObjectSize =
797           M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType());
798       BitsPtr = &Bits.back();
799     }
800 
801     for (MDNode *Type : Types) {
802       auto TypeID = Type->getOperand(1).get();
803 
804       uint64_t Offset =
805           cast<ConstantInt>(
806               cast<ConstantAsMetadata>(Type->getOperand(0))->getValue())
807               ->getZExtValue();
808 
809       TypeIdMap[TypeID].insert({BitsPtr, Offset});
810     }
811   }
812 }
813 
814 bool DevirtModule::tryFindVirtualCallTargets(
815     std::vector<VirtualCallTarget> &TargetsForSlot,
816     const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset) {
817   for (const TypeMemberInfo &TM : TypeMemberInfos) {
818     if (!TM.Bits->GV->isConstant())
819       return false;
820 
821     Constant *Ptr = getPointerAtOffset(TM.Bits->GV->getInitializer(),
822                                        TM.Offset + ByteOffset, M);
823     if (!Ptr)
824       return false;
825 
826     auto Fn = dyn_cast<Function>(Ptr->stripPointerCasts());
827     if (!Fn)
828       return false;
829 
830     // We can disregard __cxa_pure_virtual as a possible call target, as
831     // calls to pure virtuals are UB.
832     if (Fn->getName() == "__cxa_pure_virtual")
833       continue;
834 
835     TargetsForSlot.push_back({Fn, &TM});
836   }
837 
838   // Give up if we couldn't find any targets.
839   return !TargetsForSlot.empty();
840 }
841 
842 bool DevirtIndex::tryFindVirtualCallTargets(
843     std::vector<ValueInfo> &TargetsForSlot, const TypeIdCompatibleVtableInfo TIdInfo,
844     uint64_t ByteOffset) {
845   for (const TypeIdOffsetVtableInfo &P : TIdInfo) {
846     // Ensure that we have at most one external linkage vtable initializer.
847     assert(P.VTableVI.getSummaryList().size() == 1 ||
848            llvm::count_if(
849                P.VTableVI.getSummaryList(),
850                [&](const std::unique_ptr<GlobalValueSummary> &Summary) {
851                  return GlobalValue::isExternalLinkage(Summary->linkage());
852                }) <= 1);
853     // Find the first non-available_externally linkage vtable initializer.
854     // We can have multiple available_externally, linkonce_odr and weak_odr
855     // vtable initializers, however we want to skip available_externally as they
856     // do not have type metadata attached, and therefore the summary will not
857     // contain any vtable functions.
858     //
859     // Also, handle the case of same-named local Vtables with the same path
860     // and therefore the same GUID. This can happen if there isn't enough
861     // distinguishing path when compiling the source file. In that case we
862     // conservatively return false early.
863     const GlobalVarSummary *VS = nullptr;
864     bool LocalFound = false;
865     for (auto &S : P.VTableVI.getSummaryList()) {
866       if (GlobalValue::isLocalLinkage(S->linkage())) {
867         if (LocalFound)
868           return false;
869         LocalFound = true;
870       }
871       if (!GlobalValue::isAvailableExternallyLinkage(S->linkage()))
872         VS = cast<GlobalVarSummary>(S.get());
873     }
874     if (!VS->isLive())
875       continue;
876     for (auto VTP : VS->vTableFuncs()) {
877       if (VTP.VTableOffset != P.AddressPointOffset + ByteOffset)
878         continue;
879 
880       TargetsForSlot.push_back(VTP.FuncVI);
881     }
882   }
883 
884   // Give up if we couldn't find any targets.
885   return !TargetsForSlot.empty();
886 }
887 
888 void DevirtModule::applySingleImplDevirt(VTableSlotInfo &SlotInfo,
889                                          Constant *TheFn, bool &IsExported) {
890   auto Apply = [&](CallSiteInfo &CSInfo) {
891     for (auto &&VCallSite : CSInfo.CallSites) {
892       if (RemarksEnabled)
893         VCallSite.emitRemark("single-impl",
894                              TheFn->stripPointerCasts()->getName(), OREGetter);
895       VCallSite.CS.setCalledFunction(ConstantExpr::getBitCast(
896           TheFn, VCallSite.CS.getCalledValue()->getType()));
897       // This use is no longer unsafe.
898       if (VCallSite.NumUnsafeUses)
899         --*VCallSite.NumUnsafeUses;
900     }
901     if (CSInfo.isExported())
902       IsExported = true;
903     CSInfo.markDevirt();
904   };
905   Apply(SlotInfo.CSInfo);
906   for (auto &P : SlotInfo.ConstCSInfo)
907     Apply(P.second);
908 }
909 
910 static bool AddCalls(VTableSlotInfo &SlotInfo, const ValueInfo &Callee) {
911   // We can't add calls if we haven't seen a definition
912   if (Callee.getSummaryList().empty())
913     return false;
914 
915   // Insert calls into the summary index so that the devirtualized targets
916   // are eligible for import.
917   // FIXME: Annotate type tests with hotness. For now, mark these as hot
918   // to better ensure we have the opportunity to inline them.
919   bool IsExported = false;
920   auto &S = Callee.getSummaryList()[0];
921   CalleeInfo CI(CalleeInfo::HotnessType::Hot, /* RelBF = */ 0);
922   auto AddCalls = [&](CallSiteInfo &CSInfo) {
923     for (auto *FS : CSInfo.SummaryTypeCheckedLoadUsers) {
924       FS->addCall({Callee, CI});
925       IsExported |= S->modulePath() != FS->modulePath();
926     }
927     for (auto *FS : CSInfo.SummaryTypeTestAssumeUsers) {
928       FS->addCall({Callee, CI});
929       IsExported |= S->modulePath() != FS->modulePath();
930     }
931   };
932   AddCalls(SlotInfo.CSInfo);
933   for (auto &P : SlotInfo.ConstCSInfo)
934     AddCalls(P.second);
935   return IsExported;
936 }
937 
938 bool DevirtModule::trySingleImplDevirt(
939     ModuleSummaryIndex *ExportSummary,
940     MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo,
941     WholeProgramDevirtResolution *Res) {
942   // See if the program contains a single implementation of this virtual
943   // function.
944   Function *TheFn = TargetsForSlot[0].Fn;
945   for (auto &&Target : TargetsForSlot)
946     if (TheFn != Target.Fn)
947       return false;
948 
949   // If so, update each call site to call that implementation directly.
950   if (RemarksEnabled)
951     TargetsForSlot[0].WasDevirt = true;
952 
953   bool IsExported = false;
954   applySingleImplDevirt(SlotInfo, TheFn, IsExported);
955   if (!IsExported)
956     return false;
957 
958   // If the only implementation has local linkage, we must promote to external
959   // to make it visible to thin LTO objects. We can only get here during the
960   // ThinLTO export phase.
961   if (TheFn->hasLocalLinkage()) {
962     std::string NewName = (TheFn->getName() + "$merged").str();
963 
964     // Since we are renaming the function, any comdats with the same name must
965     // also be renamed. This is required when targeting COFF, as the comdat name
966     // must match one of the names of the symbols in the comdat.
967     if (Comdat *C = TheFn->getComdat()) {
968       if (C->getName() == TheFn->getName()) {
969         Comdat *NewC = M.getOrInsertComdat(NewName);
970         NewC->setSelectionKind(C->getSelectionKind());
971         for (GlobalObject &GO : M.global_objects())
972           if (GO.getComdat() == C)
973             GO.setComdat(NewC);
974       }
975     }
976 
977     TheFn->setLinkage(GlobalValue::ExternalLinkage);
978     TheFn->setVisibility(GlobalValue::HiddenVisibility);
979     TheFn->setName(NewName);
980   }
981   if (ValueInfo TheFnVI = ExportSummary->getValueInfo(TheFn->getGUID()))
982     // Any needed promotion of 'TheFn' has already been done during
983     // LTO unit split, so we can ignore return value of AddCalls.
984     AddCalls(SlotInfo, TheFnVI);
985 
986   Res->TheKind = WholeProgramDevirtResolution::SingleImpl;
987   Res->SingleImplName = TheFn->getName();
988 
989   return true;
990 }
991 
992 bool DevirtIndex::trySingleImplDevirt(MutableArrayRef<ValueInfo> TargetsForSlot,
993                                       VTableSlotSummary &SlotSummary,
994                                       VTableSlotInfo &SlotInfo,
995                                       WholeProgramDevirtResolution *Res,
996                                       std::set<ValueInfo> &DevirtTargets) {
997   // See if the program contains a single implementation of this virtual
998   // function.
999   auto TheFn = TargetsForSlot[0];
1000   for (auto &&Target : TargetsForSlot)
1001     if (TheFn != Target)
1002       return false;
1003 
1004   // Don't devirtualize if we don't have target definition.
1005   auto Size = TheFn.getSummaryList().size();
1006   if (!Size)
1007     return false;
1008 
1009   // If the summary list contains multiple summaries where at least one is
1010   // a local, give up, as we won't know which (possibly promoted) name to use.
1011   for (auto &S : TheFn.getSummaryList())
1012     if (GlobalValue::isLocalLinkage(S->linkage()) && Size > 1)
1013       return false;
1014 
1015   // Collect functions devirtualized at least for one call site for stats.
1016   if (PrintSummaryDevirt)
1017     DevirtTargets.insert(TheFn);
1018 
1019   auto &S = TheFn.getSummaryList()[0];
1020   bool IsExported = AddCalls(SlotInfo, TheFn);
1021   if (IsExported)
1022     ExportedGUIDs.insert(TheFn.getGUID());
1023 
1024   // Record in summary for use in devirtualization during the ThinLTO import
1025   // step.
1026   Res->TheKind = WholeProgramDevirtResolution::SingleImpl;
1027   if (GlobalValue::isLocalLinkage(S->linkage())) {
1028     if (IsExported)
1029       // If target is a local function and we are exporting it by
1030       // devirtualizing a call in another module, we need to record the
1031       // promoted name.
1032       Res->SingleImplName = ModuleSummaryIndex::getGlobalNameForLocal(
1033           TheFn.name(), ExportSummary.getModuleHash(S->modulePath()));
1034     else {
1035       LocalWPDTargetsMap[TheFn].push_back(SlotSummary);
1036       Res->SingleImplName = TheFn.name();
1037     }
1038   } else
1039     Res->SingleImplName = TheFn.name();
1040 
1041   // Name will be empty if this thin link driven off of serialized combined
1042   // index (e.g. llvm-lto). However, WPD is not supported/invoked for the
1043   // legacy LTO API anyway.
1044   assert(!Res->SingleImplName.empty());
1045 
1046   return true;
1047 }
1048 
1049 void DevirtModule::tryICallBranchFunnel(
1050     MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo,
1051     WholeProgramDevirtResolution *Res, VTableSlot Slot) {
1052   Triple T(M.getTargetTriple());
1053   if (T.getArch() != Triple::x86_64)
1054     return;
1055 
1056   if (TargetsForSlot.size() > ClThreshold)
1057     return;
1058 
1059   bool HasNonDevirt = !SlotInfo.CSInfo.AllCallSitesDevirted;
1060   if (!HasNonDevirt)
1061     for (auto &P : SlotInfo.ConstCSInfo)
1062       if (!P.second.AllCallSitesDevirted) {
1063         HasNonDevirt = true;
1064         break;
1065       }
1066 
1067   if (!HasNonDevirt)
1068     return;
1069 
1070   FunctionType *FT =
1071       FunctionType::get(Type::getVoidTy(M.getContext()), {Int8PtrTy}, true);
1072   Function *JT;
1073   if (isa<MDString>(Slot.TypeID)) {
1074     JT = Function::Create(FT, Function::ExternalLinkage,
1075                           M.getDataLayout().getProgramAddressSpace(),
1076                           getGlobalName(Slot, {}, "branch_funnel"), &M);
1077     JT->setVisibility(GlobalValue::HiddenVisibility);
1078   } else {
1079     JT = Function::Create(FT, Function::InternalLinkage,
1080                           M.getDataLayout().getProgramAddressSpace(),
1081                           "branch_funnel", &M);
1082   }
1083   JT->addAttribute(1, Attribute::Nest);
1084 
1085   std::vector<Value *> JTArgs;
1086   JTArgs.push_back(JT->arg_begin());
1087   for (auto &T : TargetsForSlot) {
1088     JTArgs.push_back(getMemberAddr(T.TM));
1089     JTArgs.push_back(T.Fn);
1090   }
1091 
1092   BasicBlock *BB = BasicBlock::Create(M.getContext(), "", JT, nullptr);
1093   Function *Intr =
1094       Intrinsic::getDeclaration(&M, llvm::Intrinsic::icall_branch_funnel, {});
1095 
1096   auto *CI = CallInst::Create(Intr, JTArgs, "", BB);
1097   CI->setTailCallKind(CallInst::TCK_MustTail);
1098   ReturnInst::Create(M.getContext(), nullptr, BB);
1099 
1100   bool IsExported = false;
1101   applyICallBranchFunnel(SlotInfo, JT, IsExported);
1102   if (IsExported)
1103     Res->TheKind = WholeProgramDevirtResolution::BranchFunnel;
1104 }
1105 
1106 void DevirtModule::applyICallBranchFunnel(VTableSlotInfo &SlotInfo,
1107                                           Constant *JT, bool &IsExported) {
1108   auto Apply = [&](CallSiteInfo &CSInfo) {
1109     if (CSInfo.isExported())
1110       IsExported = true;
1111     if (CSInfo.AllCallSitesDevirted)
1112       return;
1113     for (auto &&VCallSite : CSInfo.CallSites) {
1114       CallSite CS = VCallSite.CS;
1115 
1116       // Jump tables are only profitable if the retpoline mitigation is enabled.
1117       Attribute FSAttr = CS.getCaller()->getFnAttribute("target-features");
1118       if (FSAttr.hasAttribute(Attribute::None) ||
1119           !FSAttr.getValueAsString().contains("+retpoline"))
1120         continue;
1121 
1122       if (RemarksEnabled)
1123         VCallSite.emitRemark("branch-funnel",
1124                              JT->stripPointerCasts()->getName(), OREGetter);
1125 
1126       // Pass the address of the vtable in the nest register, which is r10 on
1127       // x86_64.
1128       std::vector<Type *> NewArgs;
1129       NewArgs.push_back(Int8PtrTy);
1130       for (Type *T : CS.getFunctionType()->params())
1131         NewArgs.push_back(T);
1132       FunctionType *NewFT =
1133           FunctionType::get(CS.getFunctionType()->getReturnType(), NewArgs,
1134                             CS.getFunctionType()->isVarArg());
1135       PointerType *NewFTPtr = PointerType::getUnqual(NewFT);
1136 
1137       IRBuilder<> IRB(CS.getInstruction());
1138       std::vector<Value *> Args;
1139       Args.push_back(IRB.CreateBitCast(VCallSite.VTable, Int8PtrTy));
1140       for (unsigned I = 0; I != CS.getNumArgOperands(); ++I)
1141         Args.push_back(CS.getArgOperand(I));
1142 
1143       CallSite NewCS;
1144       if (CS.isCall())
1145         NewCS = IRB.CreateCall(NewFT, IRB.CreateBitCast(JT, NewFTPtr), Args);
1146       else
1147         NewCS = IRB.CreateInvoke(
1148             NewFT, IRB.CreateBitCast(JT, NewFTPtr),
1149             cast<InvokeInst>(CS.getInstruction())->getNormalDest(),
1150             cast<InvokeInst>(CS.getInstruction())->getUnwindDest(), Args);
1151       NewCS.setCallingConv(CS.getCallingConv());
1152 
1153       AttributeList Attrs = CS.getAttributes();
1154       std::vector<AttributeSet> NewArgAttrs;
1155       NewArgAttrs.push_back(AttributeSet::get(
1156           M.getContext(), ArrayRef<Attribute>{Attribute::get(
1157                               M.getContext(), Attribute::Nest)}));
1158       for (unsigned I = 0; I + 2 <  Attrs.getNumAttrSets(); ++I)
1159         NewArgAttrs.push_back(Attrs.getParamAttributes(I));
1160       NewCS.setAttributes(
1161           AttributeList::get(M.getContext(), Attrs.getFnAttributes(),
1162                              Attrs.getRetAttributes(), NewArgAttrs));
1163 
1164       CS->replaceAllUsesWith(NewCS.getInstruction());
1165       CS->eraseFromParent();
1166 
1167       // This use is no longer unsafe.
1168       if (VCallSite.NumUnsafeUses)
1169         --*VCallSite.NumUnsafeUses;
1170     }
1171     // Don't mark as devirtualized because there may be callers compiled without
1172     // retpoline mitigation, which would mean that they are lowered to
1173     // llvm.type.test and therefore require an llvm.type.test resolution for the
1174     // type identifier.
1175   };
1176   Apply(SlotInfo.CSInfo);
1177   for (auto &P : SlotInfo.ConstCSInfo)
1178     Apply(P.second);
1179 }
1180 
1181 bool DevirtModule::tryEvaluateFunctionsWithArgs(
1182     MutableArrayRef<VirtualCallTarget> TargetsForSlot,
1183     ArrayRef<uint64_t> Args) {
1184   // Evaluate each function and store the result in each target's RetVal
1185   // field.
1186   for (VirtualCallTarget &Target : TargetsForSlot) {
1187     if (Target.Fn->arg_size() != Args.size() + 1)
1188       return false;
1189 
1190     Evaluator Eval(M.getDataLayout(), nullptr);
1191     SmallVector<Constant *, 2> EvalArgs;
1192     EvalArgs.push_back(
1193         Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0)));
1194     for (unsigned I = 0; I != Args.size(); ++I) {
1195       auto *ArgTy = dyn_cast<IntegerType>(
1196           Target.Fn->getFunctionType()->getParamType(I + 1));
1197       if (!ArgTy)
1198         return false;
1199       EvalArgs.push_back(ConstantInt::get(ArgTy, Args[I]));
1200     }
1201 
1202     Constant *RetVal;
1203     if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) ||
1204         !isa<ConstantInt>(RetVal))
1205       return false;
1206     Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue();
1207   }
1208   return true;
1209 }
1210 
1211 void DevirtModule::applyUniformRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
1212                                          uint64_t TheRetVal) {
1213   for (auto Call : CSInfo.CallSites)
1214     Call.replaceAndErase(
1215         "uniform-ret-val", FnName, RemarksEnabled, OREGetter,
1216         ConstantInt::get(cast<IntegerType>(Call.CS.getType()), TheRetVal));
1217   CSInfo.markDevirt();
1218 }
1219 
1220 bool DevirtModule::tryUniformRetValOpt(
1221     MutableArrayRef<VirtualCallTarget> TargetsForSlot, CallSiteInfo &CSInfo,
1222     WholeProgramDevirtResolution::ByArg *Res) {
1223   // Uniform return value optimization. If all functions return the same
1224   // constant, replace all calls with that constant.
1225   uint64_t TheRetVal = TargetsForSlot[0].RetVal;
1226   for (const VirtualCallTarget &Target : TargetsForSlot)
1227     if (Target.RetVal != TheRetVal)
1228       return false;
1229 
1230   if (CSInfo.isExported()) {
1231     Res->TheKind = WholeProgramDevirtResolution::ByArg::UniformRetVal;
1232     Res->Info = TheRetVal;
1233   }
1234 
1235   applyUniformRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), TheRetVal);
1236   if (RemarksEnabled)
1237     for (auto &&Target : TargetsForSlot)
1238       Target.WasDevirt = true;
1239   return true;
1240 }
1241 
1242 std::string DevirtModule::getGlobalName(VTableSlot Slot,
1243                                         ArrayRef<uint64_t> Args,
1244                                         StringRef Name) {
1245   std::string FullName = "__typeid_";
1246   raw_string_ostream OS(FullName);
1247   OS << cast<MDString>(Slot.TypeID)->getString() << '_' << Slot.ByteOffset;
1248   for (uint64_t Arg : Args)
1249     OS << '_' << Arg;
1250   OS << '_' << Name;
1251   return OS.str();
1252 }
1253 
1254 bool DevirtModule::shouldExportConstantsAsAbsoluteSymbols() {
1255   Triple T(M.getTargetTriple());
1256   return (T.getArch() == Triple::x86 || T.getArch() == Triple::x86_64) &&
1257          T.getObjectFormat() == Triple::ELF;
1258 }
1259 
1260 void DevirtModule::exportGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
1261                                 StringRef Name, Constant *C) {
1262   GlobalAlias *GA = GlobalAlias::create(Int8Ty, 0, GlobalValue::ExternalLinkage,
1263                                         getGlobalName(Slot, Args, Name), C, &M);
1264   GA->setVisibility(GlobalValue::HiddenVisibility);
1265 }
1266 
1267 void DevirtModule::exportConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
1268                                   StringRef Name, uint32_t Const,
1269                                   uint32_t &Storage) {
1270   if (shouldExportConstantsAsAbsoluteSymbols()) {
1271     exportGlobal(
1272         Slot, Args, Name,
1273         ConstantExpr::getIntToPtr(ConstantInt::get(Int32Ty, Const), Int8PtrTy));
1274     return;
1275   }
1276 
1277   Storage = Const;
1278 }
1279 
1280 Constant *DevirtModule::importGlobal(VTableSlot Slot, ArrayRef<uint64_t> Args,
1281                                      StringRef Name) {
1282   Constant *C = M.getOrInsertGlobal(getGlobalName(Slot, Args, Name), Int8Ty);
1283   auto *GV = dyn_cast<GlobalVariable>(C);
1284   if (GV)
1285     GV->setVisibility(GlobalValue::HiddenVisibility);
1286   return C;
1287 }
1288 
1289 Constant *DevirtModule::importConstant(VTableSlot Slot, ArrayRef<uint64_t> Args,
1290                                        StringRef Name, IntegerType *IntTy,
1291                                        uint32_t Storage) {
1292   if (!shouldExportConstantsAsAbsoluteSymbols())
1293     return ConstantInt::get(IntTy, Storage);
1294 
1295   Constant *C = importGlobal(Slot, Args, Name);
1296   auto *GV = cast<GlobalVariable>(C->stripPointerCasts());
1297   C = ConstantExpr::getPtrToInt(C, IntTy);
1298 
1299   // We only need to set metadata if the global is newly created, in which
1300   // case it would not have hidden visibility.
1301   if (GV->hasMetadata(LLVMContext::MD_absolute_symbol))
1302     return C;
1303 
1304   auto SetAbsRange = [&](uint64_t Min, uint64_t Max) {
1305     auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Min));
1306     auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntPtrTy, Max));
1307     GV->setMetadata(LLVMContext::MD_absolute_symbol,
1308                     MDNode::get(M.getContext(), {MinC, MaxC}));
1309   };
1310   unsigned AbsWidth = IntTy->getBitWidth();
1311   if (AbsWidth == IntPtrTy->getBitWidth())
1312     SetAbsRange(~0ull, ~0ull); // Full set.
1313   else
1314     SetAbsRange(0, 1ull << AbsWidth);
1315   return C;
1316 }
1317 
1318 void DevirtModule::applyUniqueRetValOpt(CallSiteInfo &CSInfo, StringRef FnName,
1319                                         bool IsOne,
1320                                         Constant *UniqueMemberAddr) {
1321   for (auto &&Call : CSInfo.CallSites) {
1322     IRBuilder<> B(Call.CS.getInstruction());
1323     Value *Cmp =
1324         B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
1325                      B.CreateBitCast(Call.VTable, Int8PtrTy), UniqueMemberAddr);
1326     Cmp = B.CreateZExt(Cmp, Call.CS->getType());
1327     Call.replaceAndErase("unique-ret-val", FnName, RemarksEnabled, OREGetter,
1328                          Cmp);
1329   }
1330   CSInfo.markDevirt();
1331 }
1332 
1333 Constant *DevirtModule::getMemberAddr(const TypeMemberInfo *M) {
1334   Constant *C = ConstantExpr::getBitCast(M->Bits->GV, Int8PtrTy);
1335   return ConstantExpr::getGetElementPtr(Int8Ty, C,
1336                                         ConstantInt::get(Int64Ty, M->Offset));
1337 }
1338 
1339 bool DevirtModule::tryUniqueRetValOpt(
1340     unsigned BitWidth, MutableArrayRef<VirtualCallTarget> TargetsForSlot,
1341     CallSiteInfo &CSInfo, WholeProgramDevirtResolution::ByArg *Res,
1342     VTableSlot Slot, ArrayRef<uint64_t> Args) {
1343   // IsOne controls whether we look for a 0 or a 1.
1344   auto tryUniqueRetValOptFor = [&](bool IsOne) {
1345     const TypeMemberInfo *UniqueMember = nullptr;
1346     for (const VirtualCallTarget &Target : TargetsForSlot) {
1347       if (Target.RetVal == (IsOne ? 1 : 0)) {
1348         if (UniqueMember)
1349           return false;
1350         UniqueMember = Target.TM;
1351       }
1352     }
1353 
1354     // We should have found a unique member or bailed out by now. We already
1355     // checked for a uniform return value in tryUniformRetValOpt.
1356     assert(UniqueMember);
1357 
1358     Constant *UniqueMemberAddr = getMemberAddr(UniqueMember);
1359     if (CSInfo.isExported()) {
1360       Res->TheKind = WholeProgramDevirtResolution::ByArg::UniqueRetVal;
1361       Res->Info = IsOne;
1362 
1363       exportGlobal(Slot, Args, "unique_member", UniqueMemberAddr);
1364     }
1365 
1366     // Replace each call with the comparison.
1367     applyUniqueRetValOpt(CSInfo, TargetsForSlot[0].Fn->getName(), IsOne,
1368                          UniqueMemberAddr);
1369 
1370     // Update devirtualization statistics for targets.
1371     if (RemarksEnabled)
1372       for (auto &&Target : TargetsForSlot)
1373         Target.WasDevirt = true;
1374 
1375     return true;
1376   };
1377 
1378   if (BitWidth == 1) {
1379     if (tryUniqueRetValOptFor(true))
1380       return true;
1381     if (tryUniqueRetValOptFor(false))
1382       return true;
1383   }
1384   return false;
1385 }
1386 
1387 void DevirtModule::applyVirtualConstProp(CallSiteInfo &CSInfo, StringRef FnName,
1388                                          Constant *Byte, Constant *Bit) {
1389   for (auto Call : CSInfo.CallSites) {
1390     auto *RetType = cast<IntegerType>(Call.CS.getType());
1391     IRBuilder<> B(Call.CS.getInstruction());
1392     Value *Addr =
1393         B.CreateGEP(Int8Ty, B.CreateBitCast(Call.VTable, Int8PtrTy), Byte);
1394     if (RetType->getBitWidth() == 1) {
1395       Value *Bits = B.CreateLoad(Int8Ty, Addr);
1396       Value *BitsAndBit = B.CreateAnd(Bits, Bit);
1397       auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0));
1398       Call.replaceAndErase("virtual-const-prop-1-bit", FnName, RemarksEnabled,
1399                            OREGetter, IsBitSet);
1400     } else {
1401       Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo());
1402       Value *Val = B.CreateLoad(RetType, ValAddr);
1403       Call.replaceAndErase("virtual-const-prop", FnName, RemarksEnabled,
1404                            OREGetter, Val);
1405     }
1406   }
1407   CSInfo.markDevirt();
1408 }
1409 
1410 bool DevirtModule::tryVirtualConstProp(
1411     MutableArrayRef<VirtualCallTarget> TargetsForSlot, VTableSlotInfo &SlotInfo,
1412     WholeProgramDevirtResolution *Res, VTableSlot Slot) {
1413   // This only works if the function returns an integer.
1414   auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType());
1415   if (!RetType)
1416     return false;
1417   unsigned BitWidth = RetType->getBitWidth();
1418   if (BitWidth > 64)
1419     return false;
1420 
1421   // Make sure that each function is defined, does not access memory, takes at
1422   // least one argument, does not use its first argument (which we assume is
1423   // 'this'), and has the same return type.
1424   //
1425   // Note that we test whether this copy of the function is readnone, rather
1426   // than testing function attributes, which must hold for any copy of the
1427   // function, even a less optimized version substituted at link time. This is
1428   // sound because the virtual constant propagation optimizations effectively
1429   // inline all implementations of the virtual function into each call site,
1430   // rather than using function attributes to perform local optimization.
1431   for (VirtualCallTarget &Target : TargetsForSlot) {
1432     if (Target.Fn->isDeclaration() ||
1433         computeFunctionBodyMemoryAccess(*Target.Fn, AARGetter(*Target.Fn)) !=
1434             MAK_ReadNone ||
1435         Target.Fn->arg_empty() || !Target.Fn->arg_begin()->use_empty() ||
1436         Target.Fn->getReturnType() != RetType)
1437       return false;
1438   }
1439 
1440   for (auto &&CSByConstantArg : SlotInfo.ConstCSInfo) {
1441     if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first))
1442       continue;
1443 
1444     WholeProgramDevirtResolution::ByArg *ResByArg = nullptr;
1445     if (Res)
1446       ResByArg = &Res->ResByArg[CSByConstantArg.first];
1447 
1448     if (tryUniformRetValOpt(TargetsForSlot, CSByConstantArg.second, ResByArg))
1449       continue;
1450 
1451     if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second,
1452                            ResByArg, Slot, CSByConstantArg.first))
1453       continue;
1454 
1455     // Find an allocation offset in bits in all vtables associated with the
1456     // type.
1457     uint64_t AllocBefore =
1458         findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth);
1459     uint64_t AllocAfter =
1460         findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth);
1461 
1462     // Calculate the total amount of padding needed to store a value at both
1463     // ends of the object.
1464     uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0;
1465     for (auto &&Target : TargetsForSlot) {
1466       TotalPaddingBefore += std::max<int64_t>(
1467           (AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0);
1468       TotalPaddingAfter += std::max<int64_t>(
1469           (AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0);
1470     }
1471 
1472     // If the amount of padding is too large, give up.
1473     // FIXME: do something smarter here.
1474     if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128)
1475       continue;
1476 
1477     // Calculate the offset to the value as a (possibly negative) byte offset
1478     // and (if applicable) a bit offset, and store the values in the targets.
1479     int64_t OffsetByte;
1480     uint64_t OffsetBit;
1481     if (TotalPaddingBefore <= TotalPaddingAfter)
1482       setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte,
1483                             OffsetBit);
1484     else
1485       setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte,
1486                            OffsetBit);
1487 
1488     if (RemarksEnabled)
1489       for (auto &&Target : TargetsForSlot)
1490         Target.WasDevirt = true;
1491 
1492 
1493     if (CSByConstantArg.second.isExported()) {
1494       ResByArg->TheKind = WholeProgramDevirtResolution::ByArg::VirtualConstProp;
1495       exportConstant(Slot, CSByConstantArg.first, "byte", OffsetByte,
1496                      ResByArg->Byte);
1497       exportConstant(Slot, CSByConstantArg.first, "bit", 1ULL << OffsetBit,
1498                      ResByArg->Bit);
1499     }
1500 
1501     // Rewrite each call to a load from OffsetByte/OffsetBit.
1502     Constant *ByteConst = ConstantInt::get(Int32Ty, OffsetByte);
1503     Constant *BitConst = ConstantInt::get(Int8Ty, 1ULL << OffsetBit);
1504     applyVirtualConstProp(CSByConstantArg.second,
1505                           TargetsForSlot[0].Fn->getName(), ByteConst, BitConst);
1506   }
1507   return true;
1508 }
1509 
1510 void DevirtModule::rebuildGlobal(VTableBits &B) {
1511   if (B.Before.Bytes.empty() && B.After.Bytes.empty())
1512     return;
1513 
1514   // Align the before byte array to the global's minimum alignment so that we
1515   // don't break any alignment requirements on the global.
1516   MaybeAlign Alignment(B.GV->getAlignment());
1517   if (!Alignment)
1518     Alignment =
1519         Align(M.getDataLayout().getABITypeAlignment(B.GV->getValueType()));
1520   B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), Alignment));
1521 
1522   // Before was stored in reverse order; flip it now.
1523   for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I)
1524     std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]);
1525 
1526   // Build an anonymous global containing the before bytes, followed by the
1527   // original initializer, followed by the after bytes.
1528   auto NewInit = ConstantStruct::getAnon(
1529       {ConstantDataArray::get(M.getContext(), B.Before.Bytes),
1530        B.GV->getInitializer(),
1531        ConstantDataArray::get(M.getContext(), B.After.Bytes)});
1532   auto NewGV =
1533       new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(),
1534                          GlobalVariable::PrivateLinkage, NewInit, "", B.GV);
1535   NewGV->setSection(B.GV->getSection());
1536   NewGV->setComdat(B.GV->getComdat());
1537   NewGV->setAlignment(MaybeAlign(B.GV->getAlignment()));
1538 
1539   // Copy the original vtable's metadata to the anonymous global, adjusting
1540   // offsets as required.
1541   NewGV->copyMetadata(B.GV, B.Before.Bytes.size());
1542 
1543   // Build an alias named after the original global, pointing at the second
1544   // element (the original initializer).
1545   auto Alias = GlobalAlias::create(
1546       B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "",
1547       ConstantExpr::getGetElementPtr(
1548           NewInit->getType(), NewGV,
1549           ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0),
1550                                ConstantInt::get(Int32Ty, 1)}),
1551       &M);
1552   Alias->setVisibility(B.GV->getVisibility());
1553   Alias->takeName(B.GV);
1554 
1555   B.GV->replaceAllUsesWith(Alias);
1556   B.GV->eraseFromParent();
1557 }
1558 
1559 bool DevirtModule::areRemarksEnabled() {
1560   const auto &FL = M.getFunctionList();
1561   for (const Function &Fn : FL) {
1562     const auto &BBL = Fn.getBasicBlockList();
1563     if (BBL.empty())
1564       continue;
1565     auto DI = OptimizationRemark(DEBUG_TYPE, "", DebugLoc(), &BBL.front());
1566     return DI.isEnabled();
1567   }
1568   return false;
1569 }
1570 
1571 void DevirtModule::scanTypeTestUsers(Function *TypeTestFunc) {
1572   // Find all virtual calls via a virtual table pointer %p under an assumption
1573   // of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p
1574   // points to a member of the type identifier %md. Group calls by (type ID,
1575   // offset) pair (effectively the identity of the virtual function) and store
1576   // to CallSlots.
1577   DenseSet<CallSite> SeenCallSites;
1578   for (auto I = TypeTestFunc->use_begin(), E = TypeTestFunc->use_end();
1579        I != E;) {
1580     auto CI = dyn_cast<CallInst>(I->getUser());
1581     ++I;
1582     if (!CI)
1583       continue;
1584 
1585     // Search for virtual calls based on %p and add them to DevirtCalls.
1586     SmallVector<DevirtCallSite, 1> DevirtCalls;
1587     SmallVector<CallInst *, 1> Assumes;
1588     auto &DT = LookupDomTree(*CI->getFunction());
1589     findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT);
1590 
1591     // If we found any, add them to CallSlots.
1592     if (!Assumes.empty()) {
1593       Metadata *TypeId =
1594           cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata();
1595       Value *Ptr = CI->getArgOperand(0)->stripPointerCasts();
1596       for (DevirtCallSite Call : DevirtCalls) {
1597         // Only add this CallSite if we haven't seen it before. The vtable
1598         // pointer may have been CSE'd with pointers from other call sites,
1599         // and we don't want to process call sites multiple times. We can't
1600         // just skip the vtable Ptr if it has been seen before, however, since
1601         // it may be shared by type tests that dominate different calls.
1602         if (SeenCallSites.insert(Call.CS).second)
1603           CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS, nullptr);
1604       }
1605     }
1606 
1607     // We no longer need the assumes or the type test.
1608     for (auto Assume : Assumes)
1609       Assume->eraseFromParent();
1610     // We can't use RecursivelyDeleteTriviallyDeadInstructions here because we
1611     // may use the vtable argument later.
1612     if (CI->use_empty())
1613       CI->eraseFromParent();
1614   }
1615 }
1616 
1617 void DevirtModule::scanTypeCheckedLoadUsers(Function *TypeCheckedLoadFunc) {
1618   Function *TypeTestFunc = Intrinsic::getDeclaration(&M, Intrinsic::type_test);
1619 
1620   for (auto I = TypeCheckedLoadFunc->use_begin(),
1621             E = TypeCheckedLoadFunc->use_end();
1622        I != E;) {
1623     auto CI = dyn_cast<CallInst>(I->getUser());
1624     ++I;
1625     if (!CI)
1626       continue;
1627 
1628     Value *Ptr = CI->getArgOperand(0);
1629     Value *Offset = CI->getArgOperand(1);
1630     Value *TypeIdValue = CI->getArgOperand(2);
1631     Metadata *TypeId = cast<MetadataAsValue>(TypeIdValue)->getMetadata();
1632 
1633     SmallVector<DevirtCallSite, 1> DevirtCalls;
1634     SmallVector<Instruction *, 1> LoadedPtrs;
1635     SmallVector<Instruction *, 1> Preds;
1636     bool HasNonCallUses = false;
1637     auto &DT = LookupDomTree(*CI->getFunction());
1638     findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds,
1639                                                HasNonCallUses, CI, DT);
1640 
1641     // Start by generating "pessimistic" code that explicitly loads the function
1642     // pointer from the vtable and performs the type check. If possible, we will
1643     // eliminate the load and the type check later.
1644 
1645     // If possible, only generate the load at the point where it is used.
1646     // This helps avoid unnecessary spills.
1647     IRBuilder<> LoadB(
1648         (LoadedPtrs.size() == 1 && !HasNonCallUses) ? LoadedPtrs[0] : CI);
1649     Value *GEP = LoadB.CreateGEP(Int8Ty, Ptr, Offset);
1650     Value *GEPPtr = LoadB.CreateBitCast(GEP, PointerType::getUnqual(Int8PtrTy));
1651     Value *LoadedValue = LoadB.CreateLoad(Int8PtrTy, GEPPtr);
1652 
1653     for (Instruction *LoadedPtr : LoadedPtrs) {
1654       LoadedPtr->replaceAllUsesWith(LoadedValue);
1655       LoadedPtr->eraseFromParent();
1656     }
1657 
1658     // Likewise for the type test.
1659     IRBuilder<> CallB((Preds.size() == 1 && !HasNonCallUses) ? Preds[0] : CI);
1660     CallInst *TypeTestCall = CallB.CreateCall(TypeTestFunc, {Ptr, TypeIdValue});
1661 
1662     for (Instruction *Pred : Preds) {
1663       Pred->replaceAllUsesWith(TypeTestCall);
1664       Pred->eraseFromParent();
1665     }
1666 
1667     // We have already erased any extractvalue instructions that refer to the
1668     // intrinsic call, but the intrinsic may have other non-extractvalue uses
1669     // (although this is unlikely). In that case, explicitly build a pair and
1670     // RAUW it.
1671     if (!CI->use_empty()) {
1672       Value *Pair = UndefValue::get(CI->getType());
1673       IRBuilder<> B(CI);
1674       Pair = B.CreateInsertValue(Pair, LoadedValue, {0});
1675       Pair = B.CreateInsertValue(Pair, TypeTestCall, {1});
1676       CI->replaceAllUsesWith(Pair);
1677     }
1678 
1679     // The number of unsafe uses is initially the number of uses.
1680     auto &NumUnsafeUses = NumUnsafeUsesForTypeTest[TypeTestCall];
1681     NumUnsafeUses = DevirtCalls.size();
1682 
1683     // If the function pointer has a non-call user, we cannot eliminate the type
1684     // check, as one of those users may eventually call the pointer. Increment
1685     // the unsafe use count to make sure it cannot reach zero.
1686     if (HasNonCallUses)
1687       ++NumUnsafeUses;
1688     for (DevirtCallSite Call : DevirtCalls) {
1689       CallSlots[{TypeId, Call.Offset}].addCallSite(Ptr, Call.CS,
1690                                                    &NumUnsafeUses);
1691     }
1692 
1693     CI->eraseFromParent();
1694   }
1695 }
1696 
1697 void DevirtModule::importResolution(VTableSlot Slot, VTableSlotInfo &SlotInfo) {
1698   auto *TypeId = dyn_cast<MDString>(Slot.TypeID);
1699   if (!TypeId)
1700     return;
1701   const TypeIdSummary *TidSummary =
1702       ImportSummary->getTypeIdSummary(TypeId->getString());
1703   if (!TidSummary)
1704     return;
1705   auto ResI = TidSummary->WPDRes.find(Slot.ByteOffset);
1706   if (ResI == TidSummary->WPDRes.end())
1707     return;
1708   const WholeProgramDevirtResolution &Res = ResI->second;
1709 
1710   if (Res.TheKind == WholeProgramDevirtResolution::SingleImpl) {
1711     assert(!Res.SingleImplName.empty());
1712     // The type of the function in the declaration is irrelevant because every
1713     // call site will cast it to the correct type.
1714     Constant *SingleImpl =
1715         cast<Constant>(M.getOrInsertFunction(Res.SingleImplName,
1716                                              Type::getVoidTy(M.getContext()))
1717                            .getCallee());
1718 
1719     // This is the import phase so we should not be exporting anything.
1720     bool IsExported = false;
1721     applySingleImplDevirt(SlotInfo, SingleImpl, IsExported);
1722     assert(!IsExported);
1723   }
1724 
1725   for (auto &CSByConstantArg : SlotInfo.ConstCSInfo) {
1726     auto I = Res.ResByArg.find(CSByConstantArg.first);
1727     if (I == Res.ResByArg.end())
1728       continue;
1729     auto &ResByArg = I->second;
1730     // FIXME: We should figure out what to do about the "function name" argument
1731     // to the apply* functions, as the function names are unavailable during the
1732     // importing phase. For now we just pass the empty string. This does not
1733     // impact correctness because the function names are just used for remarks.
1734     switch (ResByArg.TheKind) {
1735     case WholeProgramDevirtResolution::ByArg::UniformRetVal:
1736       applyUniformRetValOpt(CSByConstantArg.second, "", ResByArg.Info);
1737       break;
1738     case WholeProgramDevirtResolution::ByArg::UniqueRetVal: {
1739       Constant *UniqueMemberAddr =
1740           importGlobal(Slot, CSByConstantArg.first, "unique_member");
1741       applyUniqueRetValOpt(CSByConstantArg.second, "", ResByArg.Info,
1742                            UniqueMemberAddr);
1743       break;
1744     }
1745     case WholeProgramDevirtResolution::ByArg::VirtualConstProp: {
1746       Constant *Byte = importConstant(Slot, CSByConstantArg.first, "byte",
1747                                       Int32Ty, ResByArg.Byte);
1748       Constant *Bit = importConstant(Slot, CSByConstantArg.first, "bit", Int8Ty,
1749                                      ResByArg.Bit);
1750       applyVirtualConstProp(CSByConstantArg.second, "", Byte, Bit);
1751       break;
1752     }
1753     default:
1754       break;
1755     }
1756   }
1757 
1758   if (Res.TheKind == WholeProgramDevirtResolution::BranchFunnel) {
1759     // The type of the function is irrelevant, because it's bitcast at calls
1760     // anyhow.
1761     Constant *JT = cast<Constant>(
1762         M.getOrInsertFunction(getGlobalName(Slot, {}, "branch_funnel"),
1763                               Type::getVoidTy(M.getContext()))
1764             .getCallee());
1765     bool IsExported = false;
1766     applyICallBranchFunnel(SlotInfo, JT, IsExported);
1767     assert(!IsExported);
1768   }
1769 }
1770 
1771 void DevirtModule::removeRedundantTypeTests() {
1772   auto True = ConstantInt::getTrue(M.getContext());
1773   for (auto &&U : NumUnsafeUsesForTypeTest) {
1774     if (U.second == 0) {
1775       U.first->replaceAllUsesWith(True);
1776       U.first->eraseFromParent();
1777     }
1778   }
1779 }
1780 
1781 bool DevirtModule::run() {
1782   // If only some of the modules were split, we cannot correctly perform
1783   // this transformation. We already checked for the presense of type tests
1784   // with partially split modules during the thin link, and would have emitted
1785   // an error if any were found, so here we can simply return.
1786   if ((ExportSummary && ExportSummary->partiallySplitLTOUnits()) ||
1787       (ImportSummary && ImportSummary->partiallySplitLTOUnits()))
1788     return false;
1789 
1790   Function *TypeTestFunc =
1791       M.getFunction(Intrinsic::getName(Intrinsic::type_test));
1792   Function *TypeCheckedLoadFunc =
1793       M.getFunction(Intrinsic::getName(Intrinsic::type_checked_load));
1794   Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume));
1795 
1796   // Normally if there are no users of the devirtualization intrinsics in the
1797   // module, this pass has nothing to do. But if we are exporting, we also need
1798   // to handle any users that appear only in the function summaries.
1799   if (!ExportSummary &&
1800       (!TypeTestFunc || TypeTestFunc->use_empty() || !AssumeFunc ||
1801        AssumeFunc->use_empty()) &&
1802       (!TypeCheckedLoadFunc || TypeCheckedLoadFunc->use_empty()))
1803     return false;
1804 
1805   if (TypeTestFunc && AssumeFunc)
1806     scanTypeTestUsers(TypeTestFunc);
1807 
1808   if (TypeCheckedLoadFunc)
1809     scanTypeCheckedLoadUsers(TypeCheckedLoadFunc);
1810 
1811   if (ImportSummary) {
1812     for (auto &S : CallSlots)
1813       importResolution(S.first, S.second);
1814 
1815     removeRedundantTypeTests();
1816 
1817     // The rest of the code is only necessary when exporting or during regular
1818     // LTO, so we are done.
1819     return true;
1820   }
1821 
1822   // Rebuild type metadata into a map for easy lookup.
1823   std::vector<VTableBits> Bits;
1824   DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap;
1825   buildTypeIdentifierMap(Bits, TypeIdMap);
1826   if (TypeIdMap.empty())
1827     return true;
1828 
1829   // Collect information from summary about which calls to try to devirtualize.
1830   if (ExportSummary) {
1831     DenseMap<GlobalValue::GUID, TinyPtrVector<Metadata *>> MetadataByGUID;
1832     for (auto &P : TypeIdMap) {
1833       if (auto *TypeId = dyn_cast<MDString>(P.first))
1834         MetadataByGUID[GlobalValue::getGUID(TypeId->getString())].push_back(
1835             TypeId);
1836     }
1837 
1838     for (auto &P : *ExportSummary) {
1839       for (auto &S : P.second.SummaryList) {
1840         auto *FS = dyn_cast<FunctionSummary>(S.get());
1841         if (!FS)
1842           continue;
1843         // FIXME: Only add live functions.
1844         for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) {
1845           for (Metadata *MD : MetadataByGUID[VF.GUID]) {
1846             CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS);
1847           }
1848         }
1849         for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) {
1850           for (Metadata *MD : MetadataByGUID[VF.GUID]) {
1851             CallSlots[{MD, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS);
1852           }
1853         }
1854         for (const FunctionSummary::ConstVCall &VC :
1855              FS->type_test_assume_const_vcalls()) {
1856           for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) {
1857             CallSlots[{MD, VC.VFunc.Offset}]
1858                 .ConstCSInfo[VC.Args]
1859                 .addSummaryTypeTestAssumeUser(FS);
1860           }
1861         }
1862         for (const FunctionSummary::ConstVCall &VC :
1863              FS->type_checked_load_const_vcalls()) {
1864           for (Metadata *MD : MetadataByGUID[VC.VFunc.GUID]) {
1865             CallSlots[{MD, VC.VFunc.Offset}]
1866                 .ConstCSInfo[VC.Args]
1867                 .addSummaryTypeCheckedLoadUser(FS);
1868           }
1869         }
1870       }
1871     }
1872   }
1873 
1874   // For each (type, offset) pair:
1875   bool DidVirtualConstProp = false;
1876   std::map<std::string, Function*> DevirtTargets;
1877   for (auto &S : CallSlots) {
1878     // Search each of the members of the type identifier for the virtual
1879     // function implementation at offset S.first.ByteOffset, and add to
1880     // TargetsForSlot.
1881     std::vector<VirtualCallTarget> TargetsForSlot;
1882     if (tryFindVirtualCallTargets(TargetsForSlot, TypeIdMap[S.first.TypeID],
1883                                   S.first.ByteOffset)) {
1884       WholeProgramDevirtResolution *Res = nullptr;
1885       if (ExportSummary && isa<MDString>(S.first.TypeID))
1886         Res = &ExportSummary
1887                    ->getOrInsertTypeIdSummary(
1888                        cast<MDString>(S.first.TypeID)->getString())
1889                    .WPDRes[S.first.ByteOffset];
1890 
1891       if (!trySingleImplDevirt(ExportSummary, TargetsForSlot, S.second, Res)) {
1892         DidVirtualConstProp |=
1893             tryVirtualConstProp(TargetsForSlot, S.second, Res, S.first);
1894 
1895         tryICallBranchFunnel(TargetsForSlot, S.second, Res, S.first);
1896       }
1897 
1898       // Collect functions devirtualized at least for one call site for stats.
1899       if (RemarksEnabled)
1900         for (const auto &T : TargetsForSlot)
1901           if (T.WasDevirt)
1902             DevirtTargets[T.Fn->getName()] = T.Fn;
1903     }
1904 
1905     // CFI-specific: if we are exporting and any llvm.type.checked.load
1906     // intrinsics were *not* devirtualized, we need to add the resulting
1907     // llvm.type.test intrinsics to the function summaries so that the
1908     // LowerTypeTests pass will export them.
1909     if (ExportSummary && isa<MDString>(S.first.TypeID)) {
1910       auto GUID =
1911           GlobalValue::getGUID(cast<MDString>(S.first.TypeID)->getString());
1912       for (auto FS : S.second.CSInfo.SummaryTypeCheckedLoadUsers)
1913         FS->addTypeTest(GUID);
1914       for (auto &CCS : S.second.ConstCSInfo)
1915         for (auto FS : CCS.second.SummaryTypeCheckedLoadUsers)
1916           FS->addTypeTest(GUID);
1917     }
1918   }
1919 
1920   if (RemarksEnabled) {
1921     // Generate remarks for each devirtualized function.
1922     for (const auto &DT : DevirtTargets) {
1923       Function *F = DT.second;
1924 
1925       using namespace ore;
1926       OREGetter(F).emit(OptimizationRemark(DEBUG_TYPE, "Devirtualized", F)
1927                         << "devirtualized "
1928                         << NV("FunctionName", DT.first));
1929     }
1930   }
1931 
1932   removeRedundantTypeTests();
1933 
1934   // Rebuild each global we touched as part of virtual constant propagation to
1935   // include the before and after bytes.
1936   if (DidVirtualConstProp)
1937     for (VTableBits &B : Bits)
1938       rebuildGlobal(B);
1939 
1940   // We have lowered or deleted the type checked load intrinsics, so we no
1941   // longer have enough information to reason about the liveness of virtual
1942   // function pointers in GlobalDCE.
1943   for (GlobalVariable &GV : M.globals())
1944     GV.eraseMetadata(LLVMContext::MD_vcall_visibility);
1945 
1946   return true;
1947 }
1948 
1949 void DevirtIndex::run() {
1950   if (ExportSummary.typeIdCompatibleVtableMap().empty())
1951     return;
1952 
1953   DenseMap<GlobalValue::GUID, std::vector<StringRef>> NameByGUID;
1954   for (auto &P : ExportSummary.typeIdCompatibleVtableMap()) {
1955     NameByGUID[GlobalValue::getGUID(P.first)].push_back(P.first);
1956   }
1957 
1958   // Collect information from summary about which calls to try to devirtualize.
1959   for (auto &P : ExportSummary) {
1960     for (auto &S : P.second.SummaryList) {
1961       auto *FS = dyn_cast<FunctionSummary>(S.get());
1962       if (!FS)
1963         continue;
1964       // FIXME: Only add live functions.
1965       for (FunctionSummary::VFuncId VF : FS->type_test_assume_vcalls()) {
1966         for (StringRef Name : NameByGUID[VF.GUID]) {
1967           CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeTestAssumeUser(FS);
1968         }
1969       }
1970       for (FunctionSummary::VFuncId VF : FS->type_checked_load_vcalls()) {
1971         for (StringRef Name : NameByGUID[VF.GUID]) {
1972           CallSlots[{Name, VF.Offset}].CSInfo.addSummaryTypeCheckedLoadUser(FS);
1973         }
1974       }
1975       for (const FunctionSummary::ConstVCall &VC :
1976            FS->type_test_assume_const_vcalls()) {
1977         for (StringRef Name : NameByGUID[VC.VFunc.GUID]) {
1978           CallSlots[{Name, VC.VFunc.Offset}]
1979               .ConstCSInfo[VC.Args]
1980               .addSummaryTypeTestAssumeUser(FS);
1981         }
1982       }
1983       for (const FunctionSummary::ConstVCall &VC :
1984            FS->type_checked_load_const_vcalls()) {
1985         for (StringRef Name : NameByGUID[VC.VFunc.GUID]) {
1986           CallSlots[{Name, VC.VFunc.Offset}]
1987               .ConstCSInfo[VC.Args]
1988               .addSummaryTypeCheckedLoadUser(FS);
1989         }
1990       }
1991     }
1992   }
1993 
1994   std::set<ValueInfo> DevirtTargets;
1995   // For each (type, offset) pair:
1996   for (auto &S : CallSlots) {
1997     // Search each of the members of the type identifier for the virtual
1998     // function implementation at offset S.first.ByteOffset, and add to
1999     // TargetsForSlot.
2000     std::vector<ValueInfo> TargetsForSlot;
2001     auto TidSummary = ExportSummary.getTypeIdCompatibleVtableSummary(S.first.TypeID);
2002     assert(TidSummary);
2003     if (tryFindVirtualCallTargets(TargetsForSlot, *TidSummary,
2004                                   S.first.ByteOffset)) {
2005       WholeProgramDevirtResolution *Res =
2006           &ExportSummary.getOrInsertTypeIdSummary(S.first.TypeID)
2007                .WPDRes[S.first.ByteOffset];
2008 
2009       if (!trySingleImplDevirt(TargetsForSlot, S.first, S.second, Res,
2010                                DevirtTargets))
2011         continue;
2012     }
2013   }
2014 
2015   // Optionally have the thin link print message for each devirtualized
2016   // function.
2017   if (PrintSummaryDevirt)
2018     for (const auto &DT : DevirtTargets)
2019       errs() << "Devirtualized call to " << DT << "\n";
2020 
2021   return;
2022 }
2023