Relative VTables ABI on Fuchsia

[leonardchan]

Summary

This contains the updated patch for enabling Peter Collingbourne's relative vtable ABI (described in https://bugs.llvm.org/show_bug.cgi?id=26723) on Fuchsia. The goal is to reduce memory usage by making vtables readonly.

Key differences:

• LTO is not required
• No white listing of classes. Fuchsia does not guarantee C++ ABI stability, so anything running on Fuchsia should either by compiled with this ABI or not depend on the C++ ABI at all.
• Instead of taking the offset between the vtable and virtual function directly, we emit a DSO-local stub that just contains a tail call to the original virtual function, and take the offset between that and the vtable. We do this because there are some cases where the original function that would've been inserted into the vtable is not local or hidden and may require some kind of dynamic relocation which prevents the vtable from being readonly. On x86_64, taking the offset between the function and the vtable gets lowered to the offset between the PLT entry for the function and the vtable which gives us a PLT32 reloc. On Aarch64, right now only CALL26 and JUMP26 instructions generate PLT relocations, so we manifest them with stubs that are just jumps to the original function.
• Apply the same technique with the RTTI component in the vtable and emit a proxy that results in a GOTPCREL reloc.
• Update libcxxabi to use different arithmetic for finding the typeinfo pointer given the vtable pointer for Fuchsia
• Hide the ABI behind a flag -fexperimental-relative-c++-abi-vtables that is only available through CCI (-Xclang will need to be passed before it)

[pcc]

On Aarch64, right now only CALL26 and JUMP26 instructions generate PLT relocations, so we manifest them with stubs that are just jumps to the original function.

I think it would be worth considering defining a new relocation type for this. I think we should make the new reloc type represent the relative address shifted right by 2, which would allow it to be used freely in the aarch64 small code model (which permits programs up to 4GB), but would require the target to be 4-byte aligned, which seems like a reasonable restriction. On aarch64, decoding this would not require any additional instructions either (we can fold the lsl into the add). We could use the same technique for a new GOTPCREL relocation to be used for the RTTI component. @peter.smith what do you think?

[peter.smith]

In D72959#1832011, @pcc wrote:

On Aarch64, right now only CALL26 and JUMP26 instructions generate PLT relocations, so we manifest them with stubs that are just jumps to the original function.

I think it would be worth considering defining a new relocation type for this. I think we should make the new reloc type represent the relative address shifted right by 2, which would allow it to be used freely in the aarch64 small code model (which permits programs up to 4GB), but would require the target to be 4-byte aligned, which seems like a reasonable restriction. On aarch64, decoding this would not require any additional instructions either (we can fold the lsl into the add). We could use the same technique for a new GOTPCREL relocation to be used for the RTTI component. @peter.smith what do you think?

The idea of a new relocation type has come up before, as I recall it was something like the equivalent of R_X86_PLT32

| R_X86_64_PLT32 | 4 | word32 | L + A - P |
Where L is defined as: L Represents the place (section offset or address) of the Procedure Linkage Table entry for a symbol.

For Fuchsia there are two options:
1.) Ask for an ABI relocation type to be defined. I've raised an ABI issue with Arm, for it to progress I think it needs a "yes we really would like one, here is the definition we want to use, this is the use case and it could be used outside of Fuchsia at some point." For example I can see position independent C++ being useful in bare-metal embedded C++. The external facing email for feedback on the abi is [e-mail] (address is on the first page of the doc below)
2.) There are two ranges of relocation types reserved for private and platform relocations: https://developer.arm.com/docs/ihi0056/f/elf-for-the-arm-64-bit-architecture-aarch64-abi-2019q2-documentation

Private and platform-specific relocations
Private relocations for vendor experiments:

0xE000 to 0xEFFF for ELF64
0xE0 to 0xEF for ELF32
Platform ABI defined relocations:

0xF000 to 0xFFFF for ELF64
0xF0 to 0xFF for ELF32
Platform ABI relocations can only be interpreted when the EI_OSABI field is set to indicate the Platform ABI governing the definition.

All of the above codes will not be assigned by any future version of this standard.

If this could be generally useful outside of Fuchsia, then an official relocation would be preferable.

[rjmccall]

There's two independently-useful things here for the relocation, right? First, it's useful to be able to express a relocation to a symbol that has the semantics of a particular function but doesn't necessarily have its address, and that concept could be used across many "physical" relocations; and second, it's potentially useful to have a shifted-by-two relative address relocation, at least on AArch64 where instructions (and v-table entries under this ABI) are always four-byte-aligned anyway. Is it possible to separate these concerns in ELF, e.g. by having a "symbol" that can be the target of any other relocation but which actually represents a function of unspecified address with the semantics of another function?

[peter.smith]

In D72959#1835368, @rjmccall wrote:

There's two independently-useful things here for the relocation, right? First, it's useful to be able to express a relocation to a symbol that has the semantics of a particular function but doesn't necessarily have its address, and that concept could be used across many "physical" relocations; and second, it's potentially useful to have a shifted-by-two relative address relocation, at least on AArch64 where instructions (and v-table entries under this ABI) are always four-byte-aligned anyway. Is it possible to separate these concerns in ELF, e.g. by having a "symbol" that can be the target of any other relocation but which actually represents a function of unspecified address with the semantics of another function?

It would be possible to design relocations like that, but I think it would be difficult to fit into existing multi-target designs, which assume that the relocation code alone encodes all the actions a linker needs to take (smart-format, dumb-linker). My understanding of the reasons behind this are:

• The linker can have millions of relocations to resolve and having all the actions explicit in the code simplifies its job.
• There is a large space of available relocation codes, partitioned per target, but a much more limited availability of target specific symbol types and flags.
• The concerns can be separated at implementation time, for example the symbol lookup, encoding and calculation stages are independent and shared between the codes.

It is possible that there is a better trade-off in complexity, but anything radically different might need quite a bit of work to fit into an existing linker. Hope I've understood you correctly and apologies if the above isn't relevant.

[rjmccall]

Okay. In that case, I would ask that if you're considering going through the work of adding relocations, please consider adding both scaled and unscaled relative-offset-to-equivalent-symbol.

[leonardchan]

I updated the patch such that this everything here allows us to work with RTTI also. Since this is very large and has a lot of refactoring of the ItaniumCXXABI class and llvm changes independent from the FuchsiaCXXABI, I'll do my best to split them into smaller patches then formally ask for reviews on them. Keeping this around for reference as the point of reference I'll keep up to date.

[ldionne]

libcxxabi/src/private_typeinfo.cpp

618

Please introduce a macro that generically expresses that relative vtables are enabled, and explain what it is. You can then enable that macro only on __Fuchsia__. I'd like to avoid freely adding platform-specific #ifdefs in the code when it's easy to avoid.

[mcgrathr]

libcxxabi/src/private_typeinfo.cpp

618

I think the libcxxabi change should be separated from the compiler change entirely.
It will need more configuration issues resolved before it usefully lands.
Let's keep this change just for the pure compiler feature.

[leonardchan]

Updated such that this ABI is only usable on Fuchsia and if -fexperimental-fuchsia-relative-c++-abi-vtables is passed.

[phosek]

clang/include/clang/Basic/LangOptions.def

355

I think we should omit the Fuchsia bit and make it a generic, this feature should be usable by any target.

clang/include/clang/Driver/Options.td

1354

Ditto, I'd drop the fuchsia bit from these flags.

[leonardchan]

• Remove fuchsia pair from the argument and LangOpt.
• Make the virtual base offset and offset to top 32-bits to make the whole vtable 32-bit aligned.
• Introduce and check against the ItaniumVTableContext::VTableComponentLayout enum which will indicate that vtable components should be relative instead of pointers. Use this instead of making overridable virtual methods in ABI classes since this shouldn't be customized between ABIs.
• Update tests.

[rjmccall]

clang/lib/CodeGen/CGVTables.cpp

623

There's already an addRelativeOffset on ConstantArrayBuilder; is that insufficient for some reason? I think that, if v-table building were refactored so that the places that build components also add them to the v-table, we'd end up with a lot more flexibility for the ABIs. We needed a similar sort of change for pointer authentication, which we haven't upstreamed to LLVM yet, but which you can see here:

https://github.com/apple/llvm-project/blob/apple/master/clang/lib/CodeGen/CGVTables.cpp

[leonardchan]

clang/lib/CodeGen/CGVTables.cpp

623

I actually did not know about this method, but it does seem to boil down to the same arithmetic used here. Will update to see if I can use the existing builders instead.

[leonardchan]

clang/lib/CodeGen/CGVTables.cpp

623

Ah, so it seems addRelativeOffset operates differently than I thought. Initially I thought it just took the offset relative to the start of the array being built, but it actually seems to be relative to the component that would be added to the builder. This is slightly different from the current implementation where the offset is instead taken relative to the address point of the current vtable.

We could technically still achieve the desired effect of offsets in the vtable if we were to use addTaggedRelativeOffset and subtract a constantaly increasing offset as more virtual function components are added, although I'm not sure how much more benefit that would offer.

An alternative approach could also be to just use addRelativeOffset for offsets relative to the component slot and update instances we access the vtable to adjust for this new arithmetic. I think this would just transform instances of llvm.load.relative(%vtable, %func_offset) to llvm.load.relative(%vtable + %func_offset , 0), which seems differerent from how it was intended to be used, but technically might work.

[rjmccall]

clang/lib/CodeGen/CGVTables.cpp

623

Interesting. Yeah, it should be relatively straightforward to add a method that adds a pointer subtraction starting from a different base, and there's already a function for getting a constant that represents the address of the current position in the struct.

I take it you've taught LLVM that it can emit that subtraction with a normal relative-offset relocation by using an addend? That must've been exciting. Or, yeah, we can add in that addend in the frontend.

I think I see the expected benefit: it should generally save an instruction from the call sequence, at least on targets like x86 that lack pre-incrementing addressing modes. It does mean that the v-table entries can't be independently described — e.g. there's really no way you could write a C++ type that behaves like a v-table entry — but that's not too much of a loss.

[leonardchan]

Updated to use builders for constructing the vtable in "leaf" functions instead of returning a value and propagating it back up to a builder. Also extended ConstantAggregateBuilderBase with a method to add a relative offset to a specific position in the aggregate being constructed.

[rjmccall]

clang/include/clang/Basic/LangOptions.def

357

"Use an ABI-incompatible v-table layout that uses relative references"

clang/include/clang/CodeGen/ConstantInitBuilder.h

230

"relative to an element in this aggregate, identified by its index".

317

Hmm, you've borrowed the last sentence from getAddrOfCurrentPosition, but it actually doesn't make any sense (in either place). I think it would be better to say something like "The returned pointer will have type T*, where T is the given type. This type can differ from the type of the actual element."

clang/lib/CodeGen/CGVTables.cpp

622

lastAddrPoint is a confusing name for this.

638

Lots of typos

646

That's not what this block is doing.

I think it would make sense to build this stub in a helper function.

Can you just avoid making the stub entirely if the function is known to be defined in this translation unit (which will include all the internal-linkage cases, but also things like hidden linkonce_odr)?

649

private is a specific linkage that you're not actually using.

653

You're propagating *all* the attributes (which is good, but the comment should be clearer).

762

Relying on a capitalization difference for this is not a good idea for readability.

836

It's not a "destructor" equivalent, it's when the virtual function is defined as deleted.

842

Prefer llvm::ConstantPointerNull::get.

911

Can we make a helper function for this?

927

Please use the same local-variable capitalization conventions as the existing code, and please don't recompute getNumVTables() each time through the loop.

I agree that vtableIndex is a clearer name than i, but if you're going to rename locals for readability, please also rename thisIndex and nextIndex.

931

"Relative offsets should be relative to the v-table address point, which is the first entry after the RTTI pointer."

Can we just do this scan ahead of time to find the component index of the address point when we're using the relative ABI? It typically won't be a large scan, and it'll be a lot more obviously correct than what you're doing here. Also, I think what you're doing here is broken for the RTTI pointer.

You could also add a method on VTableLayout to get the address point by table index.

[leonardchan]

Evidently I left a lot of spelling mistakes and leftover comments. Cleaned them up!

This should address most of @rjmccall's comments except for calculating the vtable address points beforehand which I will add soon. Just posting what I have so far.

clang/lib/CodeGen/CGVTables.cpp

646

Done and added a test for it. Was planning to add this later as a followup but it turned out to be pretty simple to implement.

927

Updated naming in addVTableComponent and addRelativeComponent also.

[leonardchan]

Calculate the address points once beforehand such that we can get an address point from a vtable index.

Also updated to emit a stub still for complete destructor components in the vtable. The reason for this is because, with optimizations enabled, it's possible for dso_local complete destructors of a child class to be replaced by destructors of a parent class in the child's vtable, even if that parent destructor is not guaranteed to be dso_local. Updated the no-stub-when-dso-local.cpp test to reflect this.

[rjmccall]

Otherwise LGTM.

clang/lib/CodeGen/CGVTables.cpp

633

componentName seems to no longer be used meaningfully. in this function.

[ldionne]

[Github PR transition cleanup]

I think this work has landed as multiple separate patches, right? If so, can we please close this to clear up the review queue?

[leonardchan]

This was landed in various other smaller changes

clang/cmake/caches/Fuchsia-stage2.cmake

if(NOT APPLE)
   set(CLANG_DEFAULT_LINKER lld CACHE STRING "")
   set(CLANG_DEFAULT_OBJCOPY llvm-objcopy CACHE STRING "")
 endif()
 set(CLANG_DEFAULT_RTLIB compiler-rt CACHE STRING "")
 set(CLANG_ENABLE_ARCMT OFF CACHE BOOL "")
 set(CLANG_ENABLE_STATIC_ANALYZER ON CACHE BOOL "")
 set(CLANG_PLUGIN_SUPPORT OFF CACHE BOOL "")
 
+set(LIBCXXABI_RELATIVE_VTABLES_ABI_ENABLED ON CACHE BOOL "")
 set(ENABLE_EXPERIMENTAL_NEW_PASS_MANAGER ON CACHE BOOL "")
 set(ENABLE_LINKER_BUILD_ID ON CACHE BOOL "")
 set(ENABLE_X86_RELAX_RELOCATIONS ON CACHE BOOL "")
 
 set(CMAKE_BUILD_TYPE Release CACHE STRING "")
 if (APPLE)
   set(MACOSX_DEPLOYMENT_TARGET 10.7 CACHE STRING "")
 endif()

 endforeach()
 
 if(FUCHSIA_SDK)
   set(FUCHSIA_aarch64_NAME arm64)
   set(FUCHSIA_i386_NAME x64)
   set(FUCHSIA_x86_64_NAME x64)
   set(FUCHSIA_riscv64_NAME riscv64)
   foreach(target i386;x86_64;aarch64;riscv64)
-    set(FUCHSIA_${target}_COMPILER_FLAGS "-I${FUCHSIA_SDK}/pkg/fdio/include")
+    set(FUCHSIA_${target}_COMPILER_FLAGS "-I${FUCHSIA_SDK}/pkg/fdio/include -Xclang -fexperimental-relative-c++-abi-vtables")
     set(FUCHSIA_${target}_LINKER_FLAGS "-L${FUCHSIA_SDK}/arch/${FUCHSIA_${target}_NAME}/lib")
     set(FUCHSIA_${target}_SYSROOT "${FUCHSIA_SDK}/arch/${FUCHSIA_${target}_NAME}/sysroot")
   endforeach()
 
   foreach(target i386;x86_64;aarch64;riscv64)
     # Set the per-target builtins options.
     list(APPEND BUILTIN_TARGETS "${target}-unknown-fuchsia")
     set(BUILTINS_${target}-unknown-fuchsia_CMAKE_SYSTEM_NAME Fuchsia CACHE STRING "")

clang/include/clang/AST/VTableBuilder.h

 public:
   typedef std::pair<uint64_t, ThunkInfo> VTableThunkTy;
   struct AddressPointLocation {
     unsigned VTableIndex, AddressPointIndex;
   };
   typedef llvm::DenseMap<BaseSubobject, AddressPointLocation>
       AddressPointsMapTy;
 
+  // Mapping between the VTable index and address point index. This is useful
+  // when you don't care about the base subobjects and only want the address
+  // point for a given vtable index.
+  typedef llvm::SmallVector<unsigned, 4> AddressPointsIndexMapTy;
+
 private:
   // Stores the component indices of the first component of each virtual table in
   // the virtual table group. To save a little memory in the common case where
   // the vtable group contains a single vtable, an empty vector here represents
   // the vector {0}.
   OwningArrayRef<size_t> VTableIndices;
 
   OwningArrayRef<VTableComponent> VTableComponents;
 
   /// Contains thunks needed by vtables, sorted by indices.
   OwningArrayRef<VTableThunkTy> VTableThunks;
 
   /// Address points for all vtables.
   AddressPointsMapTy AddressPoints;
 
+  /// Address points for all vtable indices.
+  AddressPointsIndexMapTy AddressPointIndices;
+
 public:
   VTableLayout(ArrayRef<size_t> VTableIndices,
                ArrayRef<VTableComponent> VTableComponents,
                ArrayRef<VTableThunkTy> VTableThunks,
                const AddressPointsMapTy &AddressPoints);
   ~VTableLayout();
 
   ArrayRef<VTableComponent> vtable_components() const {
     return VTableComponents;
   }
 
   ArrayRef<VTableThunkTy> vtable_thunks() const {
     return VTableThunks;
   }
 
   AddressPointLocation getAddressPoint(BaseSubobject Base) const {
     assert(AddressPoints.count(Base) && "Did not find address point!");
     return AddressPoints.find(Base)->second;
   }
 
   const AddressPointsMapTy &getAddressPoints() const {
     return AddressPoints;
   }
 
+  const AddressPointsIndexMapTy &getAddressPointIndices() const {
+    return AddressPointIndices;
+  }
+
   size_t getNumVTables() const {
     if (VTableIndices.empty())
       return 1;
     return VTableIndices.size();
   }
 
   size_t getVTableOffset(size_t i) const {
     if (VTableIndices.empty()) {
private:
   /// where the offsets for virtual bases of a class are stored.
   typedef llvm::DenseMap<ClassPairTy, CharUnits>
     VirtualBaseClassOffsetOffsetsMapTy;
   VirtualBaseClassOffsetOffsetsMapTy VirtualBaseClassOffsetOffsets;
 
   void computeVTableRelatedInformation(const CXXRecordDecl *RD) override;
 
 public:
-  ItaniumVTableContext(ASTContext &Context);
+  enum VTableComponentLayout {
+    /// Components in the vtable are pointers to other structs/functions.
+    Pointer,
+
+    /// Components in the vtable are relative offsets between the vtable and the
+    /// other structs/functions.
+    Relative,
+  };
+
+  ItaniumVTableContext(ASTContext &Context,
+                       VTableComponentLayout ComponentLayout = Pointer);
   ~ItaniumVTableContext() override;
 
   const VTableLayout &getVTableLayout(const CXXRecordDecl *RD) {
     computeVTableRelatedInformation(RD);
     assert(VTableLayouts.count(RD) && "No layout for this record decl!");
 
     return *VTableLayouts[RD];
   }
public:
   ///
   /// Base must be a virtual base class or an unambiguous base.
   CharUnits getVirtualBaseOffsetOffset(const CXXRecordDecl *RD,
                                        const CXXRecordDecl *VBase);
 
   static bool classof(const VTableContextBase *VT) {
     return !VT->isMicrosoft();
   }
+
+  VTableComponentLayout getVTableComponentLayout() const {
+    return ComponentLayout;
+  }
+
+  bool isPointerLayout() const { return ComponentLayout == Pointer; }
+  bool isRelativeLayout() const { return ComponentLayout == Relative; }
+
+private:
+  VTableComponentLayout ComponentLayout;
 };
 
 /// Holds information about the inheritance path to a virtual base or function
 /// table pointer.  A record may contain as many vfptrs or vbptrs as there are
 /// base subobjects.
 struct VPtrInfo {
   typedef SmallVector<const CXXRecordDecl *, 1> BasePath;
 

clang/include/clang/Basic/LangOptions.def

 
 LANGOPT(FixedPoint, 1, 0, "fixed point types")
 LANGOPT(PaddingOnUnsignedFixedPoint, 1, 0,
         "unsigned fixed point types having one extra padding bit")
 
 LANGOPT(RegisterStaticDestructors, 1, 1, "Register C++ static destructors")
 
 COMPATIBLE_VALUE_LANGOPT(MaxTokens, 32, 0, "Max number of tokens per TU or 0")
+LANGOPT(RelativeCXXABIVTables, 1, 0,
+        "Use an ABI-incompatible v-table layout that uses relative references")
 
 ENUM_LANGOPT(SignReturnAddressScope, SignReturnAddressScopeKind, 2, SignReturnAddressScopeKind::None,
              "Scope of return address signing")
 ENUM_LANGOPT(SignReturnAddressKey, SignReturnAddressKeyKind, 1, SignReturnAddressKeyKind::AKey,
              "Key used for return address signing")
 LANGOPT(BranchTargetEnforcement, 1, 0, "Branch-target enforcement enabled")
 
 #undef LANGOPT
 #undef COMPATIBLE_LANGOPT
 #undef BENIGN_LANGOPT
 #undef ENUM_LANGOPT
 #undef COMPATIBLE_ENUM_LANGOPT
 #undef BENIGN_ENUM_LANGOPT
 #undef VALUE_LANGOPT
 #undef COMPATIBLE_VALUE_LANGOPT
 #undef BENIGN_VALUE_LANGOPT

clang/include/clang/CodeGen/ConstantInitBuilder.h

public:
   /// of the relative offset.  The target must be known to be defined
   /// in the current linkage unit.  The offset will have the given
   /// integer type, which must be no wider than intptr_t.  Some
   /// targets may not fully support this operation.
   void addRelativeOffset(llvm::IntegerType *type, llvm::Constant *target) {
     add(getRelativeOffset(type, target));
   }
 
+  /// Same as addRelativeOffset(), but instead relative to an element in this
+  /// aggregate, identified by its index.
+  void addRelativeOffsetToPosition(llvm::IntegerType *type,
+                                   llvm::Constant *target, size_t position) {
+    add(getRelativeOffsetToPosition(type, target, position));
+  }
+
   /// Add a relative offset to the target address, plus a small
   /// constant offset.  This is primarily useful when the relative
   /// offset is known to be a multiple of (say) four and therefore
   /// the tag can be used to express an extra two bits of information.
   void addTaggedRelativeOffset(llvm::IntegerType *type,
                                llvm::Constant *address,
                                unsigned tag) {
     llvm::Constant *offset = getRelativeOffset(type, address);
void fillPlaceholder(PlaceholderPosition position, llvm::Constant *value) {
     assert(slot == nullptr && "placeholder already filled");
     slot = value;
   }
 
   /// Produce an address which will eventually point to the next
   /// position to be filled.  This is computed with an indexed
   /// getelementptr rather than by computing offsets.
   ///
-  /// The returned pointer will have type T*, where T is the given
-  /// position.
+  /// The returned pointer will have type T*, where T is the given type. This
+  /// type can differ from the type of the actual element.
   llvm::Constant *getAddrOfCurrentPosition(llvm::Type *type);
 
+  /// Produce an address which points to a position in the aggregate being
+  /// constructed. This is computed with an indexed getelementptr rather than by
+  /// computing offsets.
+  ///
+  /// The returned pointer will have type T*, where T is the given type. This
+  /// type can differ from the type of the actual element.
+  llvm::Constant *getAddrOfPosition(llvm::Type *type, size_t position);
+
   llvm::ArrayRef<llvm::Constant*> getGEPIndicesToCurrentPosition(
                            llvm::SmallVectorImpl<llvm::Constant*> &indices) {
     getGEPIndicesTo(indices, Builder.Buffer.size());
     return indices;
   }
 
 protected:
   llvm::Constant *finishArray(llvm::Type *eltTy);
   llvm::Constant *finishStruct(llvm::StructType *structTy);
 
 private:
   void getGEPIndicesTo(llvm::SmallVectorImpl<llvm::Constant*> &indices,
                        size_t position) const;
 
   llvm::Constant *getRelativeOffset(llvm::IntegerType *offsetType,
                                     llvm::Constant *target);
 
+  llvm::Constant *getRelativeOffsetToPosition(llvm::IntegerType *offsetType,
+                                              llvm::Constant *target,
+                                              size_t position);
+
   CharUnits getOffsetFromGlobalTo(size_t index) const;
 };
 
 template <class Impl, class Traits>
 class ConstantAggregateBuilderTemplateBase
     : public Traits::AggregateBuilderBase {
   using super = typename Traits::AggregateBuilderBase;
 public:

clang/include/clang/Driver/Options.td

 
 def ffine_grained_bitfield_accesses : Flag<["-"],
   "ffine-grained-bitfield-accesses">, Group<f_clang_Group>, Flags<[CC1Option]>,
   HelpText<"Use separate accesses for consecutive bitfield runs with legal widths and alignments.">;
 def fno_fine_grained_bitfield_accesses : Flag<["-"],
   "fno-fine-grained-bitfield-accesses">, Group<f_clang_Group>, Flags<[CC1Option]>,
   HelpText<"Use large-integer access for consecutive bitfield runs.">;
 
+def fexperimental_relative_cxx_abi_vtables : Flag<["-"], "fexperimental-relative-c++-abi-vtables">,
+  Group<f_Group>, Flags<[CC1Option]>,
+  HelpText<"Use the experimental C++ class ABI for classes with virtual tables">;
+def fno_experimental_relative_cxx_abi_vtables : Flag<["-"], "fno-experimental-relative-c++-abi-vtables">,
+  Group<f_Group>, Flags<[CC1Option]>,
+  HelpText<"Do not use the experimental C++ class ABI for classes with virtual tables">;
+
 def flat__namespace : Flag<["-"], "flat_namespace">;
 def flax_vector_conversions_EQ : Joined<["-"], "flax-vector-conversions=">, Group<f_Group>,
   HelpText<"Enable implicit vector bit-casts">, Values<"none,integer,all">, Flags<[CC1Option]>;
 def flax_vector_conversions : Flag<["-"], "flax-vector-conversions">, Group<f_Group>,
   Alias<flax_vector_conversions_EQ>, AliasArgs<["integer"]>;
 def flimited_precision_EQ : Joined<["-"], "flimited-precision=">, Group<f_Group>;
 def fapple_link_rtlib : Flag<["-"], "fapple-link-rtlib">, Group<f_Group>,
   HelpText<"Force linking the clang builtins runtime library">;

clang/lib/AST/ASTContext.cpp

 
 bool ASTContext::isNearlyEmpty(const CXXRecordDecl *RD) const {
   // Pass through to the C++ ABI object
   return ABI->isNearlyEmpty(RD);
 }
 
 VTableContextBase *ASTContext::getVTableContext() {
   if (!VTContext.get()) {
-    if (Target->getCXXABI().isMicrosoft())
+    auto ABI = Target->getCXXABI();
+    if (ABI.isMicrosoft())
       VTContext.reset(new MicrosoftVTableContext(*this));
-    else
-      VTContext.reset(new ItaniumVTableContext(*this));
+    else {
+      auto ComponentLayout = getLangOpts().RelativeCXXABIVTables
+                                 ? ItaniumVTableContext::Relative
+                                 : ItaniumVTableContext::Pointer;
+      VTContext.reset(new ItaniumVTableContext(*this, ComponentLayout));
+    }
   }
   return VTContext.get();
 }
 
 MangleContext *ASTContext::createMangleContext(const TargetInfo *T) {
   if (!T)
     T = Target;
   switch (T->getCXXABI().getKind()) {

clang/lib/AST/VTableBuilder.cpp

 
 /// VCallAndVBaseOffsetBuilder - Class for building vcall and vbase offsets.
 class VCallAndVBaseOffsetBuilder {
 public:
   typedef llvm::DenseMap<const CXXRecordDecl *, CharUnits>
     VBaseOffsetOffsetsMapTy;
 
 private:
+  const ItaniumVTableContext &VTables;
+
   /// MostDerivedClass - The most derived class for which we're building vcall
   /// and vbase offsets.
   const CXXRecordDecl *MostDerivedClass;
 
   /// LayoutClass - The class we're using for layout information. Will be
   /// different than the most derived class if we're building a construction
   /// vtable.
   const CXXRecordDecl *LayoutClass;
private:
   void AddVBaseOffsets(const CXXRecordDecl *Base,
                        CharUnits OffsetInLayoutClass);
 
   /// getCurrentOffsetOffset - Get the current vcall or vbase offset offset in
   /// chars, relative to the vtable address point.
   CharUnits getCurrentOffsetOffset() const;
 
 public:
-  VCallAndVBaseOffsetBuilder(const CXXRecordDecl *MostDerivedClass,
+  VCallAndVBaseOffsetBuilder(const ItaniumVTableContext &VTables,
+                             const CXXRecordDecl *MostDerivedClass,
                              const CXXRecordDecl *LayoutClass,
                              const FinalOverriders *Overriders,
                              BaseSubobject Base, bool BaseIsVirtual,
                              CharUnits OffsetInLayoutClass)
-    : MostDerivedClass(MostDerivedClass), LayoutClass(LayoutClass),
-    Context(MostDerivedClass->getASTContext()), Overriders(Overriders) {
+      : VTables(VTables), MostDerivedClass(MostDerivedClass),
+        LayoutClass(LayoutClass), Context(MostDerivedClass->getASTContext()),
+        Overriders(Overriders) {
 
     // Add vcall and vbase offsets.
     AddVCallAndVBaseOffsets(Base, BaseIsVirtual, OffsetInLayoutClass);
   }
 
   /// Methods for iterating over the components.
   typedef VTableComponentVectorTy::const_reverse_iterator const_iterator;
   const_iterator components_begin() const { return Components.rbegin(); }

 
 CharUnits VCallAndVBaseOffsetBuilder::getCurrentOffsetOffset() const {
   // OffsetIndex is the index of this vcall or vbase offset, relative to the
   // vtable address point. (We subtract 3 to account for the information just
   // above the address point, the RTTI info, the offset to top, and the
   // vcall offset itself).
   int64_t OffsetIndex = -(int64_t)(3 + Components.size());
 
-  CharUnits PointerWidth =
-    Context.toCharUnitsFromBits(Context.getTargetInfo().getPointerWidth(0));
-  CharUnits OffsetOffset = PointerWidth * OffsetIndex;
+  // Under the relative ABI, the offset widths are 32-bit ints instead of
+  // pointer widths.
+  CharUnits OffsetWidth = Context.toCharUnitsFromBits(
+      VTables.isRelativeLayout() ? 32
+                                 : Context.getTargetInfo().getPointerWidth(0));
+  CharUnits OffsetOffset = OffsetWidth * OffsetIndex;
+
   return OffsetOffset;
 }
 
 void VCallAndVBaseOffsetBuilder::AddVCallOffsets(BaseSubobject Base,
                                                  CharUnits VBaseOffset) {
   const CXXRecordDecl *RD = Base.getBase();
   const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
 
ThisAdjustment ItaniumVTableBuilder::ComputeThisAdjustment(
 
   if (Offset.VirtualBase) {
     // Get the vcall offset map for this virtual base.
     VCallOffsetMap &VCallOffsets = VCallOffsetsForVBases[Offset.VirtualBase];
 
     if (VCallOffsets.empty()) {
       // We don't have vcall offsets for this virtual base, go ahead and
       // build them.
-      VCallAndVBaseOffsetBuilder Builder(MostDerivedClass, MostDerivedClass,
-                                         /*Overriders=*/nullptr,
-                                         BaseSubobject(Offset.VirtualBase,
-                                                       CharUnits::Zero()),
-                                         /*BaseIsVirtual=*/true,
-                                         /*OffsetInLayoutClass=*/
-                                             CharUnits::Zero());
+      VCallAndVBaseOffsetBuilder Builder(
+          VTables, MostDerivedClass, MostDerivedClass,
+          BaseSubobject(Offset.VirtualBase, CharUnits::Zero()),
 
       VCallOffsets = Builder.getVCallOffsets();
     }
 
     Adjustment.Virtual.Itanium.VCallOffsetOffset =
       VCallOffsets.getVCallOffsetOffset(MD).getQuantity();
   }
 
void ItaniumVTableBuilder::LayoutPrimaryAndSecondaryVTables(
     BaseSubobject Base, bool BaseIsMorallyVirtual,
     bool BaseIsVirtualInLayoutClass, CharUnits OffsetInLayoutClass) {
   assert(Base.getBase()->isDynamicClass() && "class does not have a vtable!");
 
   unsigned VTableIndex = Components.size();
   VTableIndices.push_back(VTableIndex);
 
   // Add vcall and vbase offsets for this vtable.
-  VCallAndVBaseOffsetBuilder Builder(MostDerivedClass, LayoutClass, &Overriders,
-                                     Base, BaseIsVirtualInLayoutClass,
-                                     OffsetInLayoutClass);
+  VCallAndVBaseOffsetBuilder Builder(
+      VTables, MostDerivedClass, LayoutClass, &Overriders, Base,
+      BaseIsVirtualInLayoutClass, OffsetInLayoutClass);
   Components.append(Builder.components_begin(), Builder.components_end());
 
   // Check if we need to add these vcall offsets.
   if (BaseIsVirtualInLayoutClass && !Builder.getVCallOffsets().empty()) {
     VCallOffsetMap &VCallOffsets = VCallOffsetsForVBases[Base.getBase()];
 
     if (VCallOffsets.empty())
       VCallOffsets = Builder.getVCallOffsets();
for (const auto &I : IndicesMap) {
           << '\n';
     }
   }
 
   Out << '\n';
 }
 }
 
+static VTableLayout::AddressPointsIndexMapTy
+MakeAddressPointIndices(const VTableLayout::AddressPointsMapTy &addressPoints,
+                        unsigned numVTables) {
+  VTableLayout::AddressPointsIndexMapTy indexMap(numVTables);
+
+  for (auto it = addressPoints.begin(); it != addressPoints.end(); ++it) {
+    const auto &addressPointLoc = it->second;
+    unsigned vtableIndex = addressPointLoc.VTableIndex;
+    unsigned addressPoint = addressPointLoc.AddressPointIndex;
+    if (indexMap[vtableIndex]) {
+      // Multiple BaseSubobjects can map to the same AddressPointLocation, but
+      // every vtable index should have a unique address point.
+      assert(indexMap[vtableIndex] == addressPoint &&
+             "Every vtable index should have a unique address point. Found a "
+             "vtable that has two different address points.");
+    } else {
+      indexMap[vtableIndex] = addressPoint;
+    }
+  }
+
+  // Note that by this point, not all the address may be initialized if the
+  // AddressPoints map is empty. This is ok if the map isn't needed. See
+  // MicrosoftVTableContext::computeVTableRelatedInformation() which uses an
+  // emprt map.
+  return indexMap;
+}
+
 VTableLayout::VTableLayout(ArrayRef<size_t> VTableIndices,
                            ArrayRef<VTableComponent> VTableComponents,
                            ArrayRef<VTableThunkTy> VTableThunks,
                            const AddressPointsMapTy &AddressPoints)
     : VTableComponents(VTableComponents), VTableThunks(VTableThunks),
-      AddressPoints(AddressPoints) {
+      AddressPoints(AddressPoints), AddressPointIndices(MakeAddressPointIndices(
+                                        AddressPoints, VTableIndices.size())) {
   if (VTableIndices.size() <= 1)
     assert(VTableIndices.size() == 1 && VTableIndices[0] == 0);
   else
     this->VTableIndices = OwningArrayRef<size_t>(VTableIndices);
 
   llvm::sort(this->VTableThunks, [](const VTableLayout::VTableThunkTy &LHS,
                                     const VTableLayout::VTableThunkTy &RHS) {
     assert((LHS.first != RHS.first || LHS.second == RHS.second) &&
            "Different thunks should have unique indices!");
     return LHS.first < RHS.first;
   });
 }
 
 VTableLayout::~VTableLayout() { }
 
-ItaniumVTableContext::ItaniumVTableContext(ASTContext &Context)
-    : VTableContextBase(/*MS=*/false) {}
+ItaniumVTableContext::ItaniumVTableContext(
+    ASTContext &Context, VTableComponentLayout ComponentLayout)
+    : VTableContextBase(/*MS=*/false), ComponentLayout(ComponentLayout) {}
 
 ItaniumVTableContext::~ItaniumVTableContext() {}
 
 uint64_t ItaniumVTableContext::getMethodVTableIndex(GlobalDecl GD) {
   GD = GD.getCanonicalDecl();
   MethodVTableIndicesTy::iterator I = MethodVTableIndices.find(GD);
   if (I != MethodVTableIndices.end())
     return I->second;
ItaniumVTableContext::getVirtualBaseOffsetOffset(const CXXRecordDecl *RD,
                                                  const CXXRecordDecl *VBase) {
   ClassPairTy ClassPair(RD, VBase);
 
   VirtualBaseClassOffsetOffsetsMapTy::iterator I =
     VirtualBaseClassOffsetOffsets.find(ClassPair);
   if (I != VirtualBaseClassOffsetOffsets.end())
     return I->second;
 
-  VCallAndVBaseOffsetBuilder Builder(RD, RD, /*Overriders=*/nullptr,
+  VCallAndVBaseOffsetBuilder Builder(*this, RD, RD, /*Overriders=*/nullptr,
                                      BaseSubobject(RD, CharUnits::Zero()),
                                      /*BaseIsVirtual=*/false,
                                      /*OffsetInLayoutClass=*/CharUnits::Zero());
 
   for (const auto &I : Builder.getVBaseOffsetOffsets()) {
     // Insert all types.
     ClassPairTy ClassPair(RD, I.first);
 

clang/lib/CodeGen/CGClass.cpp

ApplyNonVirtualAndVirtualOffset(CodeGenFunction &CGF, Address addr,
                                 const CXXRecordDecl *derivedClass,
                                 const CXXRecordDecl *nearestVBase) {
   // Assert that we have something to do.
   assert(!nonVirtualOffset.isZero() || virtualOffset != nullptr);
 
   // Compute the offset from the static and dynamic components.
   llvm::Value *baseOffset;
   if (!nonVirtualOffset.isZero()) {
-    baseOffset = llvm::ConstantInt::get(CGF.PtrDiffTy,
-                                        nonVirtualOffset.getQuantity());
+    llvm::Type *OffsetType =
+        (CGF.CGM.getTarget().getCXXABI().isItaniumFamily() &&
+         CGF.CGM.getItaniumVTableContext().isRelativeLayout())
+            ? CGF.Int32Ty
+            : CGF.PtrDiffTy;
+    baseOffset =
+        llvm::ConstantInt::get(OffsetType, nonVirtualOffset.getQuantity());
     if (virtualOffset) {
       baseOffset = CGF.Builder.CreateAdd(virtualOffset, baseOffset);
     }
   } else {
     baseOffset = virtualOffset;
   }
 
   // Apply the base offset.

clang/lib/CodeGen/CGVTables.h

class CodeGenVTables {
   llvm::Constant *DeletedVirtualFn = nullptr;
 
   /// Get the address of a thunk and emit it if necessary.
   llvm::Constant *maybeEmitThunk(GlobalDecl GD,
                                  const ThunkInfo &ThunkAdjustments,
                                  bool ForVTable);
 
   void addVTableComponent(ConstantArrayBuilder &builder,
-                          const VTableLayout &layout, unsigned idx,
-                          llvm::Constant *rtti,
-                          unsigned &nextVTableThunkIndex);
+                          const VTableLayout &layout, unsigned componentIndex,
+                          llvm::Constant *rtti, unsigned &nextVTableThunkIndex,
+                          unsigned vtableAddressPoint,
+                          bool vtableHasLocalLinkage);
+
+  /// Add a 32-bit offset to a component relative to the vtable when using the
+  /// relative vtables ABI. The array builder points to the start of the vtable.
+  void addRelativeComponent(ConstantArrayBuilder &builder,
+                            llvm::Constant *component,
+                            unsigned vtableAddressPoint,
+                            bool vtableHasLocalLinkage,
+                            bool isCompleteDtor) const;
+
+  /// Create a dso_local stub that will be used for a relative reference in the
+  /// relative vtable layout. This stub will just be a tail call to the original
+  /// function and propagate any function attributes from the original. If the
+  /// original function is already dso_local, the original is returned instead
+  /// and a stub is not created.
+  llvm::Function *
+  getOrCreateRelativeStub(llvm::Function *func,
+                          llvm::GlobalValue::LinkageTypes stubLinkage,
+                          bool isCompleteDtor) const;
+
+  bool useRelativeLayout() const;
+
+  llvm::Type *getVTableComponentType() const;
 
 public:
   /// Add vtable components for the given vtable layout to the given
   /// global initializer.
   void createVTableInitializer(ConstantStructBuilder &builder,
-                               const VTableLayout &layout,
-                               llvm::Constant *rtti);
+                               const VTableLayout &layout, llvm::Constant *rtti,
+                               bool vtableHasLocalLinkage);
 
   CodeGenVTables(CodeGenModule &CGM);
 
   ItaniumVTableContext &getItaniumVTableContext() {
     return *cast<ItaniumVTableContext>(VTContext);
   }
 
   MicrosoftVTableContext &getMicrosoftVTableContext() {
public:
   void GenerateClassData(const CXXRecordDecl *RD);
 
   bool isVTableExternal(const CXXRecordDecl *RD);
 
   /// Returns the type of a vtable with the given layout. Normally a struct of
   /// arrays of pointers, with one struct element for each vtable in the vtable
   /// group.
   llvm::Type *getVTableType(const VTableLayout &layout);
+
+  /// Generate a public facing alias for the vtable and make the vtable either
+  /// hidden or private. The alias will have the original linkage and visibility
+  /// of the vtable. This is used for cases under the relative vtables ABI
+  /// when a vtable may not be dso_local.
+  void GenerateRelativeVTableAlias(llvm::GlobalVariable *VTable,
+                                   llvm::StringRef AliasName);
 };
 
 } // end namespace CodeGen
 } // end namespace clang
 #endif

clang/lib/CodeGen/CGVTables.cpp

void CodeGenVTables::EmitThunks(GlobalDecl GD) {
 
   if (!ThunkInfoVector)
     return;
 
   for (const ThunkInfo& Thunk : *ThunkInfoVector)
     maybeEmitThunk(GD, Thunk, /*ForVTable=*/false);
 }
 
-void CodeGenVTables::addVTableComponent(
-    ConstantArrayBuilder &builder, const VTableLayout &layout,
-    unsigned idx, llvm::Constant *rtti, unsigned &nextVTableThunkIndex) {
-  auto &component = layout.vtable_components()[idx];
+void CodeGenVTables::addRelativeComponent(ConstantArrayBuilder &builder,
+                                          llvm::Constant *component,
+                                          unsigned vtableAddressPoint,
+                                          bool vtableHasLocalLinkage,
+                                          bool isCompleteDtor) const {
+  // No need to get the offset of a nullptr.
+  if (component->isNullValue())
+    return builder.add(llvm::ConstantInt::get(CGM.Int32Ty, 0));
+
+  auto *globalVal =
+      cast<llvm::GlobalValue>(component->stripPointerCastsAndAliases());
+  llvm::Module &module = CGM.getModule();
+  llvm::SmallString<16> componentName(globalVal->getName());
+
+  // We don't want to copy the linkage of the vtable exactly because we still
+  // want the stub/proxy to be emitted for properly calculating the offset.
+  // Examples where there would be no symbol emitted are available_externally
+  // and private linkages.
+  auto stubLinkage = vtableHasLocalLinkage ? llvm::GlobalValue::InternalLinkage
+                                           : llvm::GlobalValue::ExternalLinkage;
+
+  llvm::Constant *target;
+  if (auto *func = dyn_cast<llvm::Function>(globalVal)) {
+    target = getOrCreateRelativeStub(func, stubLinkage, isCompleteDtor);
+  } else {
+    llvm::SmallString<16> rttiProxyName(componentName);
+    rttiProxyName.append(".rtti_proxy");
+
+    // The RTTI component may not always be emitted in the same linkage unit as
+    // the vtable. As a general case, we can make a dso_local proxy to the RTTI
+    // that points to the actual RTTI struct somewhere. This will result in a
+    // GOTPCREL relocation when taking the relative offset to the proxy.
+    llvm::GlobalVariable *proxy = module.getNamedGlobal(rttiProxyName);
+    if (!proxy) {
+      proxy = new llvm::GlobalVariable(module, globalVal->getType(),
+                                       /*isConstant=*/true, stubLinkage,
+                                       globalVal, rttiProxyName);
+      proxy->setDSOLocal(true);
+      proxy->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
+      if (!proxy->hasLocalLinkage()) {
+        proxy->setVisibility(llvm::GlobalValue::HiddenVisibility);
+        proxy->setComdat(module.getOrInsertComdat(rttiProxyName));
+      }
+    }
+    target = proxy;
+  }
+
+  builder.addRelativeOffsetToPosition(CGM.Int32Ty, target,
+                                      /*position=*/vtableAddressPoint);
+}
+
+llvm::Function *CodeGenVTables::getOrCreateRelativeStub(
+    llvm::Function *func, llvm::GlobalValue::LinkageTypes stubLinkage,
+    bool isCompleteDtor) const {
+  // A complete object destructor can later be substituted in the vtable for an
+  // appropriate base object destructor when optimizations are enabled. This can
+  // happen for child classes that don't have their own destructor. In the case
+  // where a parent virtual destructor is not guaranteed to be in the same
+  // linkage unit as the child vtable, it's possible for an external reference
+  // for this destructor to be substituted into the child vtable, preventing it
+  // from being in rodata. If this function is a complete virtual destructor, we
+  // can just force a stub to be emitted for it.
+  if (func->isDSOLocal() && !isCompleteDtor)
+    return func;
+
+  llvm::SmallString<16> stubName(func->getName());
+  stubName.append(".stub");
+
+  // Instead of taking the offset between the vtable and virtual function
+  // directly, we emit a dso_local stub that just contains a tail call to the
+  // original virtual function and take the offset between that and the
+  // vtable. We do this because there are some cases where the original
+  // function that would've been inserted into the vtable is not dso_local
+  // which may require some kind of dynamic relocation which prevents the
+  // vtable from being readonly. On x86_64, taking the offset between the
+  // function and the vtable gets lowered to the offset between the PLT entry
+  // for the function and the vtable which gives us a PLT32 reloc. On AArch64,
+  // right now only CALL26 and JUMP26 instructions generate PLT relocations,
+  // so we manifest them with stubs that are just jumps to the original
+  // function.
+  auto &module = CGM.getModule();
+  llvm::Function *stub = module.getFunction(stubName);
+  if (stub) {
+    assert(stub->isDSOLocal() &&
+           "The previous definition of this stub should've been dso_local.");
+    return stub;
+  }
+
+  stub = llvm::Function::Create(func->getFunctionType(), stubLinkage, stubName,
+                                module);
+
+  // Propogate function attributes.
+  stub->setAttributes(func->getAttributes());
+
+  stub->setDSOLocal(true);
+  stub->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
+  if (!stub->hasLocalLinkage()) {
+    stub->setVisibility(llvm::GlobalValue::HiddenVisibility);
+    stub->setComdat(module.getOrInsertComdat(stubName));
+  }
+
+  // Fill the stub with a tail call that will be optimized.
+  llvm::BasicBlock *block =
+      llvm::BasicBlock::Create(module.getContext(), "entry", stub);
+  llvm::IRBuilder<> block_builder(block);
+  llvm::SmallVector<llvm::Value *, 8> args;
+  for (auto &arg : stub->args())
+    args.push_back(&arg);
+  llvm::CallInst *call = block_builder.CreateCall(func, args);
+  call->setAttributes(func->getAttributes());
+  call->setTailCall();
+  if (call->getType()->isVoidTy())
+    block_builder.CreateRetVoid();
+  else
+    block_builder.CreateRet(call);
+
+  return stub;
+}
 
-  auto addOffsetConstant = [&](CharUnits offset) {
-    builder.add(llvm::ConstantExpr::getIntToPtr(
-        llvm::ConstantInt::get(CGM.PtrDiffTy, offset.getQuantity()),
-        CGM.Int8PtrTy));
-  };
+bool CodeGenVTables::useRelativeLayout() const {
+  return CGM.getTarget().getCXXABI().isItaniumFamily() &&
+         CGM.getItaniumVTableContext().isRelativeLayout();
+}
+
+llvm::Type *CodeGenVTables::getVTableComponentType() const {
+  if (useRelativeLayout())
+    return CGM.Int32Ty;
+  return CGM.Int8PtrTy;
+}
+
+static void AddPointerLayoutOffset(const CodeGenModule &CGM,
+                                   ConstantArrayBuilder &builder,
+                                   CharUnits offset) {
+}
+
+static void AddRelativeLayoutOffset(const CodeGenModule &CGM,
+                                    ConstantArrayBuilder &builder,
+                                    CharUnits offset) {
+  builder.add(llvm::ConstantInt::get(CGM.Int32Ty, offset.getQuantity()));
+}
+
+void CodeGenVTables::addVTableComponent(ConstantArrayBuilder &builder,
+                                        const VTableLayout &layout,
+                                        unsigned componentIndex,
+                                        llvm::Constant *rtti,
+                                        unsigned &nextVTableThunkIndex,
+                                        unsigned vtableAddressPoint,
+                                        bool vtableHasLocalLinkage) {
+  auto &component = layout.vtable_components()[componentIndex];
+
+  auto addOffsetConstant =
+      useRelativeLayout() ? AddRelativeLayoutOffset : AddPointerLayoutOffset;
 
   switch (component.getKind()) {
   case VTableComponent::CK_VCallOffset:
-    return addOffsetConstant(component.getVCallOffset());
+    return addOffsetConstant(CGM, builder, component.getVCallOffset());
 
   case VTableComponent::CK_VBaseOffset:
-    return addOffsetConstant(component.getVBaseOffset());
+    return addOffsetConstant(CGM, builder, component.getVBaseOffset());
 
   case VTableComponent::CK_OffsetToTop:
-    return addOffsetConstant(component.getOffsetToTop());
+    return addOffsetConstant(CGM, builder, component.getOffsetToTop());
 
   case VTableComponent::CK_RTTI:
+    if (useRelativeLayout())
+      return addRelativeComponent(builder, rtti, vtableAddressPoint,
+                                  vtableHasLocalLinkage,
+                                  /*isCompleteDtor=*/false);
+    else
       return builder.add(llvm::ConstantExpr::getBitCast(rtti, CGM.Int8PtrTy));
 
   case VTableComponent::CK_FunctionPointer:
   case VTableComponent::CK_CompleteDtorPointer:
   case VTableComponent::CK_DeletingDtorPointer: {
     GlobalDecl GD;
 
     // Get the right global decl.
     switch (component.getKind()) {
if (CGM.getLangOpts().CUDA) {
       const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
       // OK on device side: functions w/ __device__ attribute
       // OK on host side: anything except __device__-only functions.
       bool CanEmitMethod =
           CGM.getLangOpts().CUDAIsDevice
               ? MD->hasAttr<CUDADeviceAttr>()
               : (MD->hasAttr<CUDAHostAttr>() || !MD->hasAttr<CUDADeviceAttr>());
       if (!CanEmitMethod)
-        return builder.addNullPointer(CGM.Int8PtrTy);
+        return builder.add(llvm::ConstantExpr::getNullValue(CGM.Int8PtrTy));
       // Method is acceptable, continue processing as usual.
     }
 
     auto getSpecialVirtualFn = [&](StringRef name) -> llvm::Constant * {
+      // FIXME(PR43094): When merging comdat groups, lld can select a local
+      // symbol as the signature symbol even though it cannot be accessed
+      // outside that symbol's TU. The relative vtables ABI would make
+      // __cxa_pure_virtual and __cxa_deleted_virtual local symbols, and
+      // depending on link order, the comdat groups could resolve to the one
+      // with the local symbol. As a temporary solution, fill these components
+      // with zero. We shouldn't be calling these in the first place anyway.
+      if (useRelativeLayout())
+        return llvm::ConstantPointerNull::get(CGM.Int8PtrTy);
+
       // For NVPTX devices in OpenMP emit special functon as null pointers,
       // otherwise linking ends up with unresolved references.
       if (CGM.getLangOpts().OpenMP && CGM.getLangOpts().OpenMPIsDevice &&
           CGM.getTriple().isNVPTX())
         return llvm::ConstantPointerNull::get(CGM.Int8PtrTy);
       llvm::FunctionType *fnTy =
           llvm::FunctionType::get(CGM.VoidTy, /*isVarArg=*/false);
       llvm::Constant *fn = cast<llvm::Constant>(
           CGM.CreateRuntimeFunction(fnTy, name).getCallee());
       if (auto f = dyn_cast<llvm::Function>(fn))
         f->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
       return llvm::ConstantExpr::getBitCast(fn, CGM.Int8PtrTy);
     };
 
     llvm::Constant *fnPtr;
 
     // Pure virtual member functions.
     if (cast<CXXMethodDecl>(GD.getDecl())->isPure()) {
       if (!PureVirtualFn)
         PureVirtualFn =
             getSpecialVirtualFn(CGM.getCXXABI().GetPureVirtualCallName());
       fnPtr = PureVirtualFn;
 
     // Deleted virtual member functions.
     } else if (cast<CXXMethodDecl>(GD.getDecl())->isDeleted()) {
       if (!DeletedVirtualFn)
         DeletedVirtualFn =
             getSpecialVirtualFn(CGM.getCXXABI().GetDeletedVirtualCallName());
       fnPtr = DeletedVirtualFn;
 
     // Thunks.
     } else if (nextVTableThunkIndex < layout.vtable_thunks().size() &&
-               layout.vtable_thunks()[nextVTableThunkIndex].first == idx) {
+               layout.vtable_thunks()[nextVTableThunkIndex].first ==
+                   componentIndex) {
       auto &thunkInfo = layout.vtable_thunks()[nextVTableThunkIndex].second;
 
       nextVTableThunkIndex++;
       fnPtr = maybeEmitThunk(GD, thunkInfo, /*ForVTable=*/true);
 
     // Otherwise we can use the method definition directly.
     } else {
       llvm::Type *fnTy = CGM.getTypes().GetFunctionTypeForVTable(GD);
       fnPtr = CGM.GetAddrOfFunction(GD, fnTy, /*ForVTable=*/true);
     }
 
-    fnPtr = llvm::ConstantExpr::getBitCast(fnPtr, CGM.Int8PtrTy);
-    builder.add(fnPtr);
-    return;
+    if (useRelativeLayout()) {
+      return addRelativeComponent(
+          builder, fnPtr, vtableAddressPoint, vtableHasLocalLinkage,
+          component.getKind() == VTableComponent::CK_CompleteDtorPointer);
+    } else
+      return builder.add(llvm::ConstantExpr::getBitCast(fnPtr, CGM.Int8PtrTy));
   }
 
   case VTableComponent::CK_UnusedFunctionPointer:
+    if (useRelativeLayout())
+      return builder.add(llvm::ConstantExpr::getNullValue(CGM.Int32Ty));
+    else
       return builder.addNullPointer(CGM.Int8PtrTy);
   }
 
   llvm_unreachable("Unexpected vtable component kind");
 }
 
 llvm::Type *CodeGenVTables::getVTableType(const VTableLayout &layout) {
   SmallVector<llvm::Type *, 4> tys;
-  for (unsigned i = 0, e = layout.getNumVTables(); i != e; ++i) {
-    tys.push_back(llvm::ArrayType::get(CGM.Int8PtrTy, layout.getVTableSize(i)));
-  }
+  llvm::Type *componentType = getVTableComponentType();
+  for (unsigned i = 0, e = layout.getNumVTables(); i != e; ++i)
+    tys.push_back(llvm::ArrayType::get(componentType, layout.getVTableSize(i)));
 
   return llvm::StructType::get(CGM.getLLVMContext(), tys);
 }
 
 void CodeGenVTables::createVTableInitializer(ConstantStructBuilder &builder,
                                              const VTableLayout &layout,
-                                             llvm::Constant *rtti) {
+                                             llvm::Constant *rtti,
+                                             bool vtableHasLocalLinkage) {
+  llvm::Type *componentType = getVTableComponentType();
+
+  const auto &addressPoints = layout.getAddressPointIndices();
   unsigned nextVTableThunkIndex = 0;
-  for (unsigned i = 0, e = layout.getNumVTables(); i != e; ++i) {
-    auto vtableElem = builder.beginArray(CGM.Int8PtrTy);
-    size_t thisIndex = layout.getVTableOffset(i);
-    size_t nextIndex = thisIndex + layout.getVTableSize(i);
-    for (unsigned i = thisIndex; i != nextIndex; ++i) {
-      addVTableComponent(vtableElem, layout, i, rtti, nextVTableThunkIndex);
+  for (unsigned vtableIndex = 0, endIndex = layout.getNumVTables();
+       vtableIndex != endIndex; ++vtableIndex) {
+    auto vtableElem = builder.beginArray(componentType);
+
+    size_t vtableStart = layout.getVTableOffset(vtableIndex);
+    size_t vtableEnd = vtableStart + layout.getVTableSize(vtableIndex);
+    for (size_t componentIndex = vtableStart; componentIndex < vtableEnd;
+         ++componentIndex) {
+      addVTableComponent(vtableElem, layout, componentIndex, rtti,
+                         nextVTableThunkIndex, addressPoints[vtableIndex],
+                         vtableHasLocalLinkage);
     }
     vtableElem.finishAndAddTo(builder);
   }
 }
 
-llvm::GlobalVariable *
-CodeGenVTables::GenerateConstructionVTable(const CXXRecordDecl *RD,
-                                      const BaseSubobject &Base,
-                                      bool BaseIsVirtual,
-                                   llvm::GlobalVariable::LinkageTypes Linkage,
-                                      VTableAddressPointsMapTy& AddressPoints) {
+llvm::GlobalVariable *CodeGenVTables::GenerateConstructionVTable(
+    const CXXRecordDecl *RD, const BaseSubobject &Base, bool BaseIsVirtual,
+    VTableAddressPointsMapTy &AddressPoints) {
   if (CGDebugInfo *DI = CGM.getModuleDebugInfo())
     DI->completeClassData(Base.getBase());
 
   std::unique_ptr<VTableLayout> VTLayout(
       getItaniumVTableContext().createConstructionVTableLayout(
           Base.getBase(), Base.getBaseOffset(), BaseIsVirtual, RD));
 
   // Add the address points.
   AddressPoints = VTLayout->getAddressPoints();
 
   // Get the mangled construction vtable name.
   SmallString<256> OutName;
   llvm::raw_svector_ostream Out(OutName);
   cast<ItaniumMangleContext>(CGM.getCXXABI().getMangleContext())
       .mangleCXXCtorVTable(RD, Base.getBaseOffset().getQuantity(),
                            Base.getBase(), Out);
-  StringRef Name = OutName.str();
+  SmallString<256> Name(OutName);
+
+  bool UsingRelativeLayout = getItaniumVTableContext().isRelativeLayout();
+  bool VTableAliasExists = UsingRelativeLayout && CGM.getModule().getNamedAlias(Name);
+  if (VTableAliasExists) {
+    // We previously made the vtable hidden and changed its name.
+    Name.append(".local");
+  }
 
   llvm::Type *VTType = getVTableType(*VTLayout);
 
   // Construction vtable symbols are not part of the Itanium ABI, so we cannot
   // guarantee that they actually will be available externally. Instead, when
   // emitting an available_externally VTT, we provide references to an internal
   // linkage construction vtable. The ABI only requires complete-object vtables
   // to be the same for all instances of a type, not construction vtables.
llvm::GlobalVariable *CodeGenVTables::GenerateConstructionVTable(
   VTable->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
 
   llvm::Constant *RTTI = CGM.GetAddrOfRTTIDescriptor(
       CGM.getContext().getTagDeclType(Base.getBase()));
 
   // Create and set the initializer.
   ConstantInitBuilder builder(CGM);
   auto components = builder.beginStruct();
-  createVTableInitializer(components, *VTLayout, RTTI);
+  createVTableInitializer(components, *VTLayout, RTTI,
+                          VTable->hasLocalLinkage());
   components.finishAndSetAsInitializer(VTable);
 
   // Set properties only after the initializer has been set to ensure that the
   // GV is treated as definition and not declaration.
   assert(!VTable->isDeclaration() && "Shouldn't set properties on declaration");
   CGM.setGVProperties(VTable, RD);
 
   CGM.EmitVTableTypeMetadata(RD, VTable, *VTLayout.get());
 
+  if (UsingRelativeLayout && !VTable->isDSOLocal())
+    GenerateRelativeVTableAlias(VTable, OutName);
+
   return VTable;
 }
 
+// If the VTable is not dso_local, then we will not be able to indicate that
+// the VTable does not need a relocation and move into rodata. An frequent
+// time this can occur is for classes that should be made public from a DSO
+// (like in libc++). For cases like these, we can make the vtable hidden or
+// private and create a public alias with the same visibility and linkage as
+// the original vtable type.
+void CodeGenVTables::GenerateRelativeVTableAlias(llvm::GlobalVariable *VTable,
+                                                 llvm::StringRef AliasName) {
+  assert(getItaniumVTableContext().isRelativeLayout() &&
+         "Can only use this if the relative vtable ABI is used");
+  assert(!VTable->isDSOLocal() && "This should be called only if the vtable is "
+                                  "not guaranteed to be dso_local");
+
+  // If the vtable is available_externally, we shouldn't (or need to) generate
+  // an alias for it in the first place since the vtable won't actually by
+  // emitted in this compilation unit.
+  if (VTable->hasAvailableExternallyLinkage())
+    return;
+
+  VTable->setName(AliasName + ".local");
+
+  auto Linkage = VTable->getLinkage();
+  assert(llvm::GlobalAlias::isValidLinkage(Linkage) &&
+         "Invalid vtable alias linkage");
+
+  llvm::GlobalAlias *VTableAlias = CGM.getModule().getNamedAlias(AliasName);
+  if (!VTableAlias) {
+    VTableAlias = llvm::GlobalAlias::create(
+        VTable->getValueType(), VTable->getAddressSpace(), Linkage, AliasName,
+        &CGM.getModule());
+  } else {
+    assert(VTableAlias->getValueType() == VTable->getValueType());
+    assert(VTableAlias->getLinkage() == Linkage);
+  }
+  VTableAlias->setVisibility(VTable->getVisibility());
+  VTableAlias->setUnnamedAddr(VTable->getUnnamedAddr());
+
+  // Both of these imply dso_local for the vtable.
+  if (!VTable->hasComdat()) {
+    // If this is in a comdat, then we shouldn't make the linkage private due to
+    // an issue in lld where private symbols can be used as the key symbol when
+    // choosing the prevelant group. This leads to "relocation refers to a
+    // symbol in a discarded section".
+    VTable->setLinkage(llvm::GlobalValue::PrivateLinkage);
+  } else {
+    // We should at least make this hidden since we don't want to expose it.
+    VTable->setVisibility(llvm::GlobalValue::HiddenVisibility);
+  }
+
+  VTableAlias->setAliasee(VTable);
+}
+
 static bool shouldEmitAvailableExternallyVTable(const CodeGenModule &CGM,
                                                 const CXXRecordDecl *RD) {
   return CGM.getCodeGenOpts().OptimizationLevel > 0 &&
          CGM.getCXXABI().canSpeculativelyEmitVTable(RD);
 }
 
 /// Compute the required linkage of the vtable for the given class.
 ///

clang/lib/CodeGen/ConstantInitBuilder.cpp

 
 void ConstantAggregateBuilderBase::addSize(CharUnits size) {
   add(Builder.CGM.getSize(size));
 }
 
 llvm::Constant *
 ConstantAggregateBuilderBase::getRelativeOffset(llvm::IntegerType *offsetType,
                                                 llvm::Constant *target) {
+  return getRelativeOffsetToPosition(offsetType, target,
+                                     Builder.SelfReferences.size());
+}
+
+llvm::Constant *ConstantAggregateBuilderBase::getRelativeOffsetToPosition(
+    llvm::IntegerType *offsetType, llvm::Constant *target, size_t position) {
   // Compute the address of the relative-address slot.
-  auto base = getAddrOfCurrentPosition(offsetType);
+  auto base = getAddrOfPosition(offsetType, position);
 
   // Subtract.
   base = llvm::ConstantExpr::getPtrToInt(base, Builder.CGM.IntPtrTy);
   target = llvm::ConstantExpr::getPtrToInt(target, Builder.CGM.IntPtrTy);
   llvm::Constant *offset = llvm::ConstantExpr::getSub(target, base);
 
   // Truncate to the relative-address type if necessary.
   if (Builder.CGM.IntPtrTy != offsetType) {
     offset = llvm::ConstantExpr::getTrunc(offset, offsetType);
   }
 
   return offset;
 }
 
 llvm::Constant *
+ConstantAggregateBuilderBase::getAddrOfPosition(llvm::Type *type,
+                                                size_t position) {
+  // Make a global variable.  We will replace this with a GEP to this
+  // position after installing the initializer.
+  auto dummy = new llvm::GlobalVariable(Builder.CGM.getModule(), type, true,
+                                        llvm::GlobalVariable::PrivateLinkage,
+                                        nullptr, "");
+  Builder.SelfReferences.emplace_back(dummy);
+  auto &entry = Builder.SelfReferences.back();
+  (void)getGEPIndicesTo(entry.Indices, position + Begin);
+  return dummy;
+}
+
+llvm::Constant *
 ConstantAggregateBuilderBase::getAddrOfCurrentPosition(llvm::Type *type) {
   // Make a global variable.  We will replace this with a GEP to this
   // position after installing the initializer.
   auto dummy =
     new llvm::GlobalVariable(Builder.CGM.getModule(), type, true,
                              llvm::GlobalVariable::PrivateLinkage,
                              nullptr, "");
   Builder.SelfReferences.emplace_back(dummy);

clang/lib/CodeGen/ItaniumCXXABI.cpp

llvm::Value *VirtualFn = nullptr;
     if (ShouldEmitCFICheck || ShouldEmitVFEInfo || ShouldEmitWPDInfo) {
       // If doing CFI, VFE or WPD, we will need the metadata node to check
       // against.
       llvm::Metadata *MD =
           CGM.CreateMetadataIdentifierForVirtualMemPtrType(QualType(MPT, 0));
       TypeId = llvm::MetadataAsValue::get(CGF.getLLVMContext(), MD);
     }
 
+    if (ShouldEmitVFEInfo) {
       llvm::Value *VFPAddr = Builder.CreateGEP(VTable, VTableOffset);
 
-    if (ShouldEmitVFEInfo) {
       // If doing VFE, load from the vtable with a type.checked.load intrinsic
       // call. Note that we use the GEP to calculate the address to load from
       // and pass 0 as the offset to the intrinsic. This is because every
       // vtable slot of the correct type is marked with matching metadata, and
       // we know that the load must be from one of these slots.
       llvm::Value *CheckedLoad = Builder.CreateCall(
           CGM.getIntrinsic(llvm::Intrinsic::type_checked_load),
           {VFPAddr, llvm::ConstantInt::get(CGM.Int32Ty, 0), TypeId});
       CheckResult = Builder.CreateExtractValue(CheckedLoad, 1);
       VirtualFn = Builder.CreateExtractValue(CheckedLoad, 0);
       VirtualFn = Builder.CreateBitCast(VirtualFn, FTy->getPointerTo(),
                                         "memptr.virtualfn");
     } else {
       // When not doing VFE, emit a normal load, as it allows more
       // optimisations than type.checked.load.
       if (ShouldEmitCFICheck || ShouldEmitWPDInfo) {
+        llvm::Value *VFPAddr = Builder.CreateGEP(VTable, VTableOffset);
         CheckResult = Builder.CreateCall(
             CGM.getIntrinsic(llvm::Intrinsic::type_test),
             {Builder.CreateBitCast(VFPAddr, CGF.Int8PtrTy), TypeId});
       }
-      VFPAddr =
-          Builder.CreateBitCast(VFPAddr, FTy->getPointerTo()->getPointerTo());
-      VirtualFn = Builder.CreateAlignedLoad(VFPAddr, CGF.getPointerAlign(),
-                                            "memptr.virtualfn");
+
+      if (CGM.getItaniumVTableContext().isRelativeLayout()) {
+        VirtualFn = CGF.Builder.CreateCall(
+            CGM.getIntrinsic(llvm::Intrinsic::load_relative,
+                             {VTableOffset->getType()}),
+            {VTable, VTableOffset});
+        VirtualFn = CGF.Builder.CreateBitCast(VirtualFn, FTy->getPointerTo());
+      } else {
+        llvm::Value *VFPAddr = CGF.Builder.CreateGEP(VTable, VTableOffset);
+        VFPAddr = CGF.Builder.CreateBitCast(
+            VFPAddr, FTy->getPointerTo()->getPointerTo());
+        VirtualFn = CGF.Builder.CreateAlignedLoad(
+            VFPAddr, CGF.getPointerAlign(), "memptr.virtualfn");
+      }
     }
     assert(VirtualFn && "Virtual fuction pointer not created!");
     assert((!ShouldEmitCFICheck || !ShouldEmitVFEInfo || !ShouldEmitWPDInfo ||
             CheckResult) &&
            "Check result required but not created!");
 
     if (ShouldEmitCFICheck) {
       // If doing CFI, emit the check.
llvm::Constant *ItaniumCXXABI::BuildMemberPointer(const CXXMethodDecl *MD,
   assert(MD->isInstance() && "Member function must not be static!");
 
   CodeGenTypes &Types = CGM.getTypes();
 
   // Get the function pointer (or index if this is a virtual function).
   llvm::Constant *MemPtr[2];
   if (MD->isVirtual()) {
     uint64_t Index = CGM.getItaniumVTableContext().getMethodVTableIndex(MD);
-
+    uint64_t VTableOffset;
+    if (CGM.getItaniumVTableContext().isRelativeLayout()) {
+      // Multiply by 4-byte relative offsets.
+      VTableOffset = Index * 4;
+    } else {
       const ASTContext &Context = getContext();
-    CharUnits PointerWidth =
-      Context.toCharUnitsFromBits(Context.getTargetInfo().getPointerWidth(0));
-    uint64_t VTableOffset = (Index * PointerWidth.getQuantity());
+      CharUnits PointerWidth = Context.toCharUnitsFromBits(
+          Context.getTargetInfo().getPointerWidth(0));
+      VTableOffset = Index * PointerWidth.getQuantity();
+    }
 
     if (UseARMMethodPtrABI) {
       // ARM C++ ABI 3.2.1:
       //   This ABI specifies that adj contains twice the this
       //   adjustment, plus 1 if the member function is virtual. The
       //   least significant bit of adj then makes exactly the same
       //   discrimination as the least significant bit of ptr does for
       //   Itanium.
llvm::Value *ItaniumCXXABI::EmitTypeid(CodeGenFunction &CGF,
                                        QualType SrcRecordTy,
                                        Address ThisPtr,
                                        llvm::Type *StdTypeInfoPtrTy) {
   auto *ClassDecl =
       cast<CXXRecordDecl>(SrcRecordTy->castAs<RecordType>()->getDecl());
   llvm::Value *Value =
       CGF.GetVTablePtr(ThisPtr, StdTypeInfoPtrTy->getPointerTo(), ClassDecl);
 
+  if (CGM.getItaniumVTableContext().isRelativeLayout()) {
+    // Load the type info.
+    Value = CGF.Builder.CreateBitCast(Value, CGM.Int8PtrTy);
+    Value = CGF.Builder.CreateCall(
+        CGM.getIntrinsic(llvm::Intrinsic::load_relative, {CGM.Int32Ty}),
+        {Value, llvm::ConstantInt::get(CGM.Int32Ty, -4)});
+
+    // Setup to dereference again since this is a proxy we accessed.
+    Value = CGF.Builder.CreateBitCast(Value, StdTypeInfoPtrTy->getPointerTo());
+  } else {
     // Load the type info.
     Value = CGF.Builder.CreateConstInBoundsGEP1_64(Value, -1ULL);
+  }
   return CGF.Builder.CreateAlignedLoad(Value, CGF.getPointerAlign());
 }
 
 bool ItaniumCXXABI::shouldDynamicCastCallBeNullChecked(bool SrcIsPtr,
                                                        QualType SrcRecordTy) {
   return SrcIsPtr;
 }
 
llvm::Value *ItaniumCXXABI::EmitDynamicCastCall(
 
   return Value;
 }
 
 llvm::Value *ItaniumCXXABI::EmitDynamicCastToVoid(CodeGenFunction &CGF,
                                                   Address ThisAddr,
                                                   QualType SrcRecordTy,
                                                   QualType DestTy) {
-  llvm::Type *PtrDiffLTy =
-      CGF.ConvertType(CGF.getContext().getPointerDiffType());
   llvm::Type *DestLTy = CGF.ConvertType(DestTy);
-
   auto *ClassDecl =
       cast<CXXRecordDecl>(SrcRecordTy->castAs<RecordType>()->getDecl());
+  llvm::Value *OffsetToTop;
+  if (CGM.getItaniumVTableContext().isRelativeLayout()) {
     // Get the vtable pointer.
-  llvm::Value *VTable = CGF.GetVTablePtr(ThisAddr, PtrDiffLTy->getPointerTo(),
-      ClassDecl);
+    llvm::Value *VTable =
+        CGF.GetVTablePtr(ThisAddr, CGM.Int32Ty->getPointerTo(), ClassDecl);
 
     // Get the offset-to-top from the vtable.
-  llvm::Value *OffsetToTop =
-      CGF.Builder.CreateConstInBoundsGEP1_64(VTable, -2ULL);
     OffsetToTop =
-    CGF.Builder.CreateAlignedLoad(OffsetToTop, CGF.getPointerAlign(),
-                                  "offset.to.top");
+        CGF.Builder.CreateConstInBoundsGEP1_32(/*Type=*/nullptr, VTable, -2U);
+    OffsetToTop = CGF.Builder.CreateAlignedLoad(
+        OffsetToTop, CharUnits::fromQuantity(4), "offset.to.top");
+  } else {
+    llvm::Type *PtrDiffLTy =
+        CGF.ConvertType(CGF.getContext().getPointerDiffType());
+
+    // Get the vtable pointer.
+    llvm::Value *VTable =
+        CGF.GetVTablePtr(ThisAddr, PtrDiffLTy->getPointerTo(), ClassDecl);
 
+    // Get the offset-to-top from the vtable.
+    OffsetToTop = CGF.Builder.CreateConstInBoundsGEP1_64(VTable, -2ULL);
+    OffsetToTop = CGF.Builder.CreateAlignedLoad(
+        OffsetToTop, CGF.getPointerAlign(), "offset.to.top");
+  }
   // Finally, add the offset to the pointer.
   llvm::Value *Value = ThisAddr.getPointer();
   Value = CGF.EmitCastToVoidPtr(Value);
   Value = CGF.Builder.CreateInBoundsGEP(Value, OffsetToTop);
-
   return CGF.Builder.CreateBitCast(Value, DestLTy);
 }
 
 bool ItaniumCXXABI::EmitBadCastCall(CodeGenFunction &CGF) {
   llvm::FunctionCallee Fn = getBadCastFn(CGF);
   llvm::CallBase *Call = CGF.EmitRuntimeCallOrInvoke(Fn);
   Call->setDoesNotReturn();
   CGF.Builder.CreateUnreachable();
   return true;
 }
 
 llvm::Value *
 ItaniumCXXABI::GetVirtualBaseClassOffset(CodeGenFunction &CGF,
                                          Address This,
                                          const CXXRecordDecl *ClassDecl,
                                          const CXXRecordDecl *BaseClassDecl) {
   llvm::Value *VTablePtr = CGF.GetVTablePtr(This, CGM.Int8PtrTy, ClassDecl);
   CharUnits VBaseOffsetOffset =
       CGM.getItaniumVTableContext().getVirtualBaseOffsetOffset(ClassDecl,
                                                                BaseClassDecl);
-
   llvm::Value *VBaseOffsetPtr =
     CGF.Builder.CreateConstGEP1_64(VTablePtr, VBaseOffsetOffset.getQuantity(),
                                    "vbase.offset.ptr");
+
+  llvm::Value *VBaseOffset;
+  if (CGM.getItaniumVTableContext().isRelativeLayout()) {
+    VBaseOffsetPtr =
+        CGF.Builder.CreateBitCast(VBaseOffsetPtr, CGF.Int32Ty->getPointerTo());
+    VBaseOffset = CGF.Builder.CreateAlignedLoad(
+        VBaseOffsetPtr, CharUnits::fromQuantity(4), "vbase.offset");
+  } else {
     VBaseOffsetPtr = CGF.Builder.CreateBitCast(VBaseOffsetPtr,
                                                CGM.PtrDiffTy->getPointerTo());
-
-  llvm::Value *VBaseOffset =
-    CGF.Builder.CreateAlignedLoad(VBaseOffsetPtr, CGF.getPointerAlign(),
-                                  "vbase.offset");
-
+    VBaseOffset = CGF.Builder.CreateAlignedLoad(
+        VBaseOffsetPtr, CGF.getPointerAlign(), "vbase.offset");
+  }
   return VBaseOffset;
 }
 
 void ItaniumCXXABI::EmitCXXConstructors(const CXXConstructorDecl *D) {
   // Just make sure we're in sync with TargetCXXABI.
   assert(CGM.getTarget().getCXXABI().hasConstructorVariants());
 
   // The constructor used for constructing this as a base class;
void ItaniumCXXABI::emitVTableDefinitions(CodeGenVTables &CGVT,
 
   ItaniumVTableContext &VTContext = CGM.getItaniumVTableContext();
   const VTableLayout &VTLayout = VTContext.getVTableLayout(RD);
   llvm::GlobalVariable::LinkageTypes Linkage = CGM.getVTableLinkage(RD);
   llvm::Constant *RTTI =
       CGM.GetAddrOfRTTIDescriptor(CGM.getContext().getTagDeclType(RD));
 
   // Create and set the initializer.
-  ConstantInitBuilder Builder(CGM);
-  auto Components = Builder.beginStruct();
-  CGVT.createVTableInitializer(Components, VTLayout, RTTI);
-  Components.finishAndSetAsInitializer(VTable);
+  ConstantInitBuilder builder(CGM);
+  auto components = builder.beginStruct();
+  CGVT.createVTableInitializer(components, VTLayout, RTTI,
+                               llvm::GlobalValue::isLocalLinkage(Linkage));
+  components.finishAndSetAsInitializer(VTable);
 
   // Set the correct linkage.
   VTable->setLinkage(Linkage);
 
   if (CGM.supportsCOMDAT() && VTable->isWeakForLinker())
     VTable->setComdat(CGM.getModule().getOrInsertComdat(VTable->getName()));
 
   // Set the right visibility.
   CGM.setGVProperties(VTable, RD);
 
   // If this is the magic class __cxxabiv1::__fundamental_type_info,
   // we will emit the typeinfo for the fundamental types. This is the
   // same behaviour as GCC.
   const DeclContext *DC = RD->getDeclContext();
   if (RD->getIdentifier() &&
       RD->getIdentifier()->isStr("__fundamental_type_info") &&
       isa<NamespaceDecl>(DC) && cast<NamespaceDecl>(DC)->getIdentifier() &&
       cast<NamespaceDecl>(DC)->getIdentifier()->isStr("__cxxabiv1") &&
       DC->getParent()->isTranslationUnit())
     EmitFundamentalRTTIDescriptors(RD);
 
   if (!VTable->isDeclarationForLinker())
     CGM.EmitVTableTypeMetadata(RD, VTable, VTLayout);
+
+  if (VTContext.isRelativeLayout() && !VTable->isDSOLocal())
+    CGVT.GenerateRelativeVTableAlias(VTable, VTable->getName());
 }
 
 bool ItaniumCXXABI::isVirtualOffsetNeededForVTableField(
     CodeGenFunction &CGF, CodeGenFunction::VPtr Vptr) {
   if (Vptr.NearestVBase == nullptr)
     return false;
   return NeedsVTTParameter(CGF.CurGD);
 }
llvm::GlobalVariable *ItaniumCXXABI::getAddrOfVTable(const CXXRecordDecl *RD,
 
   const VTableLayout &VTLayout =
       CGM.getItaniumVTableContext().getVTableLayout(RD);
   llvm::Type *VTableType = CGM.getVTables().getVTableType(VTLayout);
 
   // Use pointer alignment for the vtable. Otherwise we would align them based
   // on the size of the initializer which doesn't make sense as only single
   // values are read.
-  unsigned PAlign = CGM.getTarget().getPointerAlign(0);
+  unsigned PAlign = CGM.getItaniumVTableContext().isRelativeLayout()
+                        ? 32
+                        : CGM.getTarget().getPointerAlign(0);
 
   VTable = CGM.CreateOrReplaceCXXRuntimeVariable(
       Name, VTableType, llvm::GlobalValue::ExternalLinkage,
       getContext().toCharUnitsFromBits(PAlign).getQuantity());
   VTable->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
 
   CGM.setGVProperties(VTable, RD);
 
   return VTable;
 }
 
 CGCallee ItaniumCXXABI::getVirtualFunctionPointer(CodeGenFunction &CGF,
                                                   GlobalDecl GD,
                                                   Address This,
                                                   llvm::Type *Ty,
                                                   SourceLocation Loc) {
-  Ty = Ty->getPointerTo()->getPointerTo();
   auto *MethodDecl = cast<CXXMethodDecl>(GD.getDecl());
-  llvm::Value *VTable = CGF.GetVTablePtr(This, Ty, MethodDecl->getParent());
+  llvm::Value *VTable = CGF.GetVTablePtr(
+      This, Ty->getPointerTo()->getPointerTo(), MethodDecl->getParent());
 
   uint64_t VTableIndex = CGM.getItaniumVTableContext().getMethodVTableIndex(GD);
   llvm::Value *VFunc;
   if (CGF.ShouldEmitVTableTypeCheckedLoad(MethodDecl->getParent())) {
     VFunc = CGF.EmitVTableTypeCheckedLoad(
         MethodDecl->getParent(), VTable,
         VTableIndex * CGM.getContext().getTargetInfo().getPointerWidth(0) / 8);
   } else {
     CGF.EmitTypeMetadataCodeForVCall(MethodDecl->getParent(), VTable, Loc);
 
-    llvm::Value *VFuncPtr =
+    llvm::Value *VFuncLoad;
+    if (CGM.getItaniumVTableContext().isRelativeLayout()) {
+      VTable = CGF.Builder.CreateBitCast(VTable, CGM.Int8PtrTy);
+      llvm::Value *Load = CGF.Builder.CreateCall(
+          CGM.getIntrinsic(llvm::Intrinsic::load_relative, {CGM.Int32Ty}),
+          {VTable, llvm::ConstantInt::get(CGM.Int32Ty, 4 * VTableIndex)});
+      VFuncLoad = CGF.Builder.CreateBitCast(Load, Ty->getPointerTo());
+    } else {
+      VTable =
+          CGF.Builder.CreateBitCast(VTable, Ty->getPointerTo()->getPointerTo());
+      llvm::Value *VTableSlotPtr =
           CGF.Builder.CreateConstInBoundsGEP1_64(VTable, VTableIndex, "vfn");
-    auto *VFuncLoad =
-        CGF.Builder.CreateAlignedLoad(VFuncPtr, CGF.getPointerAlign());
+      VFuncLoad =
+          CGF.Builder.CreateAlignedLoad(VTableSlotPtr, CGF.getPointerAlign());
+    }
 
     // Add !invariant.load md to virtual function load to indicate that
     // function didn't change inside vtable.
     // It's safe to add it without -fstrict-vtable-pointers, but it would not
     // help in devirtualization because it will only matter if we will have 2
     // the same virtual function loads from the same vtable load, which won't
     // happen without enabled devirtualization with -fstrict-vtable-pointers.
     if (CGM.getCodeGenOpts().OptimizationLevel > 0 &&
-        CGM.getCodeGenOpts().StrictVTablePointers)
-      VFuncLoad->setMetadata(
+        CGM.getCodeGenOpts().StrictVTablePointers) {
+      if (auto *VFuncLoadInstr = dyn_cast<llvm::Instruction>(VFuncLoad)) {
+        VFuncLoadInstr->setMetadata(
             llvm::LLVMContext::MD_invariant_load,
             llvm::MDNode::get(CGM.getLLVMContext(),
                               llvm::ArrayRef<llvm::Metadata *>()));
+      }
+    }
     VFunc = VFuncLoad;
   }
 
   CGCallee Callee(GD, VFunc);
   return Callee;
 }
 
 llvm::Value *ItaniumCXXABI::EmitVirtualDestructorCall(
static llvm::Value *performTypeAdjustment(CodeGenFunction &CGF,
   if (NonVirtualAdjustment && !IsReturnAdjustment) {
     V = CGF.Builder.CreateConstInBoundsByteGEP(V,
                               CharUnits::fromQuantity(NonVirtualAdjustment));
   }
 
   // Perform the virtual adjustment if we have one.
   llvm::Value *ResultPtr;
   if (VirtualAdjustment) {
-    llvm::Type *PtrDiffTy =
-        CGF.ConvertType(CGF.getContext().getPointerDiffType());
-
     Address VTablePtrPtr = CGF.Builder.CreateElementBitCast(V, CGF.Int8PtrTy);
     llvm::Value *VTablePtr = CGF.Builder.CreateLoad(VTablePtrPtr);
 
+    llvm::Value *Offset;
     llvm::Value *OffsetPtr =
         CGF.Builder.CreateConstInBoundsGEP1_64(VTablePtr, VirtualAdjustment);
+    if (CGF.CGM.getItaniumVTableContext().isRelativeLayout()) {
+      // Load the adjustment offset from the vtable as a 32-bit int.
+      OffsetPtr =
+          CGF.Builder.CreateBitCast(OffsetPtr, CGF.Int32Ty->getPointerTo());
+      Offset =
+          CGF.Builder.CreateAlignedLoad(OffsetPtr, CharUnits::fromQuantity(4));
+    } else {
+      llvm::Type *PtrDiffTy =
+          CGF.ConvertType(CGF.getContext().getPointerDiffType());
 
-    OffsetPtr = CGF.Builder.CreateBitCast(OffsetPtr, PtrDiffTy->getPointerTo());
+      OffsetPtr =
+          CGF.Builder.CreateBitCast(OffsetPtr, PtrDiffTy->getPointerTo());
 
       // Load the adjustment offset from the vtable.
-    llvm::Value *Offset =
-      CGF.Builder.CreateAlignedLoad(OffsetPtr, CGF.getPointerAlign());
-
+      Offset = CGF.Builder.CreateAlignedLoad(OffsetPtr, CGF.getPointerAlign());
+    }
     // Adjust our pointer.
     ResultPtr = CGF.Builder.CreateInBoundsGEP(V.getPointer(), Offset);
   } else {
     ResultPtr = V.getPointer();
   }
 
   // In a derived-to-base conversion, the non-virtual adjustment is
   // applied second.
case Type::Pointer:
     break;
 
   case Type::MemberPointer:
     // abi::__pointer_to_member_type_info.
     VTableName = "_ZTVN10__cxxabiv129__pointer_to_member_type_infoE";
     break;
   }
 
-  llvm::Constant *VTable =
-    CGM.getModule().getOrInsertGlobal(VTableName, CGM.Int8PtrTy);
+  llvm::Constant *VTable = nullptr;
+
+  // Check if the alias exists. If it doesn't, then get or create the global.
+  if (CGM.getItaniumVTableContext().isRelativeLayout())
+    VTable = CGM.getModule().getNamedAlias(VTableName);
+  if (!VTable)
+    VTable = CGM.getModule().getOrInsertGlobal(VTableName, CGM.Int8PtrTy);
+
   CGM.setDSOLocal(cast<llvm::GlobalValue>(VTable->stripPointerCasts()));
 
   llvm::Type *PtrDiffTy =
       CGM.getTypes().ConvertType(CGM.getContext().getPointerDiffType());
 
   // The vtable address point is 2.
-  llvm::Constant *Two = llvm::ConstantInt::get(PtrDiffTy, 2);
+  if (CGM.getItaniumVTableContext().isRelativeLayout()) {
+    // The vtable address point is 8 bytes after its start:
+    // 4 for the offset to top + 4 for the relative offset to rtti.
+    llvm::Constant *Eight = llvm::ConstantInt::get(CGM.Int32Ty, 8);
+    VTable = llvm::ConstantExpr::getBitCast(VTable, CGM.Int8PtrTy);
     VTable =
-      llvm::ConstantExpr::getInBoundsGetElementPtr(CGM.Int8PtrTy, VTable, Two);
+        llvm::ConstantExpr::getInBoundsGetElementPtr(CGM.Int8Ty, VTable, Eight);
+  } else {
+    llvm::Constant *Two = llvm::ConstantInt::get(PtrDiffTy, 2);
+    VTable = llvm::ConstantExpr::getInBoundsGetElementPtr(CGM.Int8PtrTy, VTable,
+                                                          Two);
+  }
   VTable = llvm::ConstantExpr::getBitCast(VTable, CGM.Int8PtrTy);
 
   Fields.push_back(VTable);
 }
 
 /// Return the linkage that the type info and type info name constants
 /// should have for the given type.
 static llvm::GlobalVariable::LinkageTypes getTypeInfoLinkage(CodeGenModule &CGM,

clang/lib/CodeGen/MicrosoftCXXABI.cpp

for (const std::unique_ptr<VPtrInfo>& Info : VFPtrs) {
     const VTableLayout &VTLayout =
       VFTContext.getVFTableLayout(RD, Info->FullOffsetInMDC);
 
     llvm::Constant *RTTI = nullptr;
     if (any_of(VTLayout.vtable_components(),
                [](const VTableComponent &VTC) { return VTC.isRTTIKind(); }))
       RTTI = getMSCompleteObjectLocator(RD, *Info);
 
-    ConstantInitBuilder Builder(CGM);
-    auto Components = Builder.beginStruct();
-    CGVT.createVTableInitializer(Components, VTLayout, RTTI);
-    Components.finishAndSetAsInitializer(VTable);
+    ConstantInitBuilder builder(CGM);
+    auto components = builder.beginStruct();
+    CGVT.createVTableInitializer(components, VTLayout, RTTI,
+                                 VTable->hasLocalLinkage());
+    components.finishAndSetAsInitializer(VTable);
 
     emitVTableTypeMetadata(*Info, RD, VTable);
   }
 }
 
 bool MicrosoftCXXABI::isVirtualOffsetNeededForVTableField(
     CodeGenFunction &CGF, CodeGenFunction::VPtr Vptr) {
   return Vptr.NearestVBase != nullptr;

clang/lib/Driver/ToolChains/Fuchsia.cpp

if (!Args.hasArg(options::OPT_nostdlib, options::OPT_nodefaultlibs)) {
 
     if (Args.hasArg(options::OPT_fsplit_stack))
       CmdArgs.push_back("--wrap=pthread_create");
 
     if (!Args.hasArg(options::OPT_nolibc))
       CmdArgs.push_back("-lc");
   }
 
+  if (Args.hasFlag(options::OPT_fexperimental_relative_cxx_abi_vtables,
+                   options::OPT_fno_experimental_relative_cxx_abi_vtables,
+                   /*default=*/false)) {
+    CmdArgs.push_back("-fexperimental-relative-c++-abi-vtables");
+  }
+
   C.addCommand(std::make_unique<Command>(JA, *this, Exec, CmdArgs, Inputs));
 }
 
 /// Fuchsia - Fuchsia tool chain which can call as(1) and ld(1) directly.
 
 Fuchsia::Fuchsia(const Driver &D, const llvm::Triple &Triple,
                  const ArgList &Args)
     : ToolChain(D, Triple, Args) {

clang/lib/Frontend/CompilerInvocation.cpp

if (Arg *A = Args.getLastArg(OPT_msign_return_address_key_EQ)) {
         else
           Diags.Report(diag::err_drv_invalid_value)
               << A->getAsString(Args) << SignKey;
       }
     }
   }
 
   Opts.BranchTargetEnforcement = Args.hasArg(OPT_mbranch_target_enforce);
+
+  Opts.RelativeCXXABIVTables =
+      Args.hasFlag(OPT_fexperimental_relative_cxx_abi_vtables,
+                   OPT_fno_experimental_relative_cxx_abi_vtables,
+                   /*default=*/false);
 }
 
 static bool isStrictlyPreprocessorAction(frontend::ActionKind Action) {
   switch (Action) {
   case frontend::ASTDeclList:
   case frontend::ASTDump:
   case frontend::ASTPrint:
   case frontend::ASTView:

clang/test/CodeGenCXX/Fuchsia/child-inheritted-from-parent-in-comdat.cpp

+// Cross comdat example
+// Parent VTable is in a comdat section.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: $_ZN1B3fooEv.stub = comdat any
+
+// The inline function is emitted in each module with the same comdat
+// CHECK: $_ZN1A3fooEv.stub = comdat any
+// CHECK: $_ZTS1A = comdat any
+// CHECK: $_ZTI1A = comdat any
+// CHECK: $_ZTI1B.rtti_proxy = comdat any
+
+// The VTable is emitted everywhere used
+// CHECK: $_ZTV1A = comdat any
+// CHECK: $_ZTI1A.rtti_proxy = comdat any
+
+// The VTable for B is emitted here since it has a key function which is defined in this module
+// CHECK: @_ZTV1B.local = private unnamed_addr constant { [3 x i32] } { [3 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i8* }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+
+// The VTable for A is emitted here and in a comdat section since it has no key function, and is used in this module when creating an instance of A (in func()).
+// CHECK: @_ZTV1A.local = linkonce_odr hidden unnamed_addr constant { [3 x i32] } { [3 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8* }** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, comdat($_ZTV1A), align 4
+
+// CHECK: @_ZTV1B = unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1B.local
+// CHECK: @_ZTV1A = linkonce_odr unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1A.local
+
+// CHECK:      define void @_ZN1B3fooEv(%class.B* nocapture %this) unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+// CHECK:      define hidden void @_ZN1B3fooEv.stub(%class.B* nocapture %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+// CHECK:      define hidden void @_ZN1A3fooEv.stub(%class.A* %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   tail call void @_ZN1A3fooEv(%class.A* %0)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  inline virtual void foo() {}
+};
+class B : public A {
+public:
+  void foo() override;
+};
+void A_foo(A *a);
+
+void B::foo() {}
+void func2() {
+  A a;
+  A_foo(&a);
+}

clang/test/CodeGenCXX/Fuchsia/child-vtable-in-comdat.cpp

+// Cross comdat example
+// Child VTable is in a comdat section.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: $_ZN1B3fooEv.stub = comdat any
+// CHECK: $_ZN1A3fooEv.stub = comdat any
+
+// A comdat is emitted for B but not A
+// CHECK: $_ZTV1B = comdat any
+// CHECK: $_ZTS1B = comdat any
+// CHECK: $_ZTI1B = comdat any
+// CHECK: $_ZTI1B.rtti_proxy = comdat any
+// CHECK: $_ZTI1A.rtti_proxy = comdat any
+
+// VTable for B is emitted here since we access it when creating an instance of B. The VTable is also linkonce_odr and in its own comdat.
+// CHECK: @_ZTV1B.local = linkonce_odr hidden unnamed_addr constant { [3 x i32] } { [3 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i8* }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, comdat($_ZTV1B), align 4
+
+// The RTTI objects aren’t that important, but it is good to know that they are emitted here since they are used in the vtable for B, and external references are used for RTTI stuff from A.
+// CHECK: @_ZTVN10__cxxabiv120__si_class_type_infoE = external global i8*
+// CHECK: @_ZTS1B = linkonce_odr constant [3 x i8] c"1B\00", comdat, align 1
+// CHECK: @_ZTI1A = external constant i8*
+// CHECK: @_ZTI1B = linkonce_odr constant { i8*, i8*, i8* } { i8* getelementptr inbounds (i8, i8* bitcast (i8** @_ZTVN10__cxxabiv120__si_class_type_infoE to i8*), i32 8), i8* getelementptr inbounds ([3 x i8], [3 x i8]* @_ZTS1B, i32 0, i32 0), i8* bitcast (i8** @_ZTI1A to i8*) }, comdat, align 8
+// CHECK: @_ZTI1B.rtti_proxy = hidden unnamed_addr constant { i8*, i8*, i8* }* @_ZTI1B, comdat
+
+// The VTable for A is available_externally, meaning it can have a definition in
+// IR, but is never emitted in this compilation unit. Because it won't be
+// emitted here, we cannot make an alias, but we won't need to in the first
+// place.
+// CHECK: @_ZTV1A = available_externally unnamed_addr constant { [3 x i32] } { [3 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint (i8*** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+// CHECK: @_ZTI1A.rtti_proxy = hidden unnamed_addr constant i8** @_ZTI1A, comdat
+
+// We will emit a vtable for B here, so it does have an alias, but we will not
+// emit one for A.
+// CHECK: @_ZTV1B = linkonce_odr unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1B.local
+// CHECK-NOT: @_ZTV1A = {{.*}}alias
+
+// CHECK:      define hidden void @_ZN1B3fooEv.stub(%class.B* %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   tail call void @_ZN1B3fooEv(%class.B* %0)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+// CHECK: declare void @_ZN1A3fooEv(%class.A*) unnamed_addr
+
+// CHECK: define hidden void @_ZN1A3fooEv.stub(%class.A* %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   tail call void @_ZN1A3fooEv(%class.A* %0)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  virtual void foo();
+};
+class B : public A {
+public:
+  inline void foo() override {}
+};
+void A_foo(A *a);
+
+// func() is used so that the vtable for B is accessed when creating the instance.
+void func() {
+  B b;
+  A_foo(&b);
+}

clang/test/CodeGenCXX/Fuchsia/cross-translation-unit-1.cpp

+// Check the vtable layout for classes with key functions defined in different
+// translation units. This TU only manifests the vtable for A.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+#include "cross-tu-header.h"
+
+// CHECK: $_ZN1A3fooEv.stub = comdat any
+// CHECK: $_ZN1A3barEv.stub = comdat any
+// CHECK: $_ZTI1A.rtti_proxy = comdat any
+
+// CHECK: @_ZTV1A.local = private unnamed_addr constant { [4 x i32] } { [4 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8* }** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3barEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+// @_ZTV1A = unnamed_addr alias { [4 x i32] }, { [4 x i32] }* @_ZTV1A.local
+
+// A::foo() is still available for other modules to use since it is not marked with private or internal linkage.
+// CHECK:      define void @_ZN1A3fooEv(%class.A* nocapture %this) unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+// The proxy that we take a reference to in the vtable has hidden visibility and external linkage so it can be used only by other modules in the same DSO. A::foo() is inlined into this stub since it is defined in the same module.
+// CHECK:      define hidden void @_ZN1A3fooEv.stub(%class.A* nocapture %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+// A::bar() is called within the module but not defined, even though the VTable for A is emitted here
+// CHECK: declare void @_ZN1A3barEv(%class.A*) unnamed_addr
+
+// The stub for A::bar() is made private, so it will not appear in the symbol table and is only used in this module. We tail call here because A::bar() is not defined in the same module.
+// CHECK:      define hidden void @_ZN1A3barEv.stub(%class.A* %0) unnamed_addr {{#[0-9]+}} comdat {
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   tail call void @_ZN1A3barEv(%class.A* %0)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+void A::foo() {}
+void A_foo(A *a) { a->foo(); }
+void A_bar(A *a) { a->bar(); }

clang/test/CodeGenCXX/Fuchsia/cross-translation-unit-2.cpp

+// Check the vtable layout for classes with key functions defined in different
+// translation units. This TU manifests the vtable for B.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+#include "cross-tu-header.h"
+
+// CHECK: $_ZN1B3fooEv.stub = comdat any
+// CHECK: $_ZN1A3barEv.stub = comdat any
+// CHECK: $_ZTI1B.rtti_proxy = comdat any
+
+// CHECK: @_ZTV1B.local = private unnamed_addr constant { [4 x i32] } { [4 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i8* }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3barEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+// CHECK: @_ZTV1B = unnamed_addr alias { [4 x i32] }, { [4 x i32] }* @_ZTV1B.local
+
+// A::bar() is defined outside of the module that defines the vtable for A
+// CHECK:      define void @_ZN1A3barEv(%class.A* nocapture %this) unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+// CHECK:      define void @_ZN1B3fooEv(%class.B* nocapture %this) unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+// The stubs for B::foo() and A::bar() are hidden
+// CHECK:      define hidden void @_ZN1B3fooEv.stub(%class.B* nocapture %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+// CHECK:      define hidden void @_ZN1A3barEv.stub(%class.A* nocapture %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+void A::bar() {}
+void B::foo() {}

clang/test/CodeGenCXX/Fuchsia/cross-tu-header.h

+class A {
+public:
+  virtual void foo();
+  virtual void bar();
+};
+
+class B : public A {
+public:
+  void foo() override;
+};

clang/test/CodeGenCXX/Fuchsia/diamond-inheritance.cpp

+// Diamond inheritance.
+// A more complicated multiple inheritance example that includes longer chain of inheritance and a common ancestor.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: %class.B = type { %class.A }
+// CHECK: %class.A = type { i32 (...)** }
+// CHECK: %class.C = type { %class.A }
+// CHECK: %class.D = type { %class.B, %class.C }
+
+// VTable for B should contain offset to top (0), RTTI pointer, A::foo(), and B::barB().
+// CHECK: @_ZTV1B.local = private unnamed_addr constant { [4 x i32] } { [4 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i8* }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B4barBEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+
+// VTable for C should contain offset to top (0), RTTI pointer, A::foo(), and C::barC().
+// CHECK: @_ZTV1C.local = private unnamed_addr constant { [4 x i32] } { [4 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i8* }** @_ZTI1C.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1C.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1C.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.C*)* @_ZN1C4barCEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1C.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+
+// VTable for D should be similar to the mutiple inheritance example where this
+// vtable contains 2 inner vtables:
+// - 1st table containing D::foo(), B::barB(), and D::baz().
+// - 2nd table containing a thunk to D::foo() and C::barC().
+// CHECK: @_ZTV1D.local = private unnamed_addr constant { [5 x i32], [4 x i32] } { [5 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64, i8*, i64 }** @_ZTI1D.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32] }, { [5 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.D*)* @_ZN1D3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32] }, { [5 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B4barBEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32] }, { [5 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.D*)* @_ZN1D3bazEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32] }, { [5 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 0, i32 2) to i64)) to i32)], [4 x i32] [i32 -8, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64, i8*, i64 }** @_ZTI1D.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32] }, { [5 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 1, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.D*)* @_ZThn8_N1D3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32] }, { [5 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 1, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.C*)* @_ZN1C4barCEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32] }, { [5 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 1, i32 2) to i64)) to i32)] }, align 4
+
+// @_ZTV1B = unnamed_addr alias { [4 x i32] }, { [4 x i32] }* @_ZTV1B.local
+// @_ZTV1C = unnamed_addr alias { [4 x i32] }, { [4 x i32] }* @_ZTV1C.local
+// @_ZTV1D = unnamed_addr alias { [5 x i32], [4 x i32] }, { [5 x i32], [4 x i32] }* @_ZTV1D.local
+
+class A {
+public:
+  virtual void foo();
+};
+
+class B : public A {
+public:
+  virtual void barB();
+};
+
+class C : public A {
+  virtual void barC();
+};
+
+// Should be a struct with 2 arrays from 2 parents.
+// The 1st contains D::foo(), B::barB(), and D::baz().
+// The 2nd contains C::barC(), and a thunk that points to D::foo().
+class D : public B, C {
+public:
+  virtual void baz();
+  void foo() override;
+};
+
+void B::barB() {}
+void C::barC() {}
+void D::foo() {}
+void D::baz() {}
+
+void D_foo(D *d) {
+  d->foo();
+}

clang/test/CodeGenCXX/Fuchsia/diamond-virtual-inheritance.cpp

+// Diamond virtual inheritance.
+// This should cover virtual inheritance, construction vtables, and VTTs.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: %class.B = type { i32 (...)**, %class.A.base }
+// CHECK: %class.A.base = type <{ i32 (...)**, i32 }>
+// CHECK: %class.C = type { i32 (...)**, %class.A.base }
+// CHECK: %class.D = type { %class.B.base, %class.C.base, %class.A.base }
+// CHECK: %class.B.base = type { i32 (...)** }
+// CHECK: %class.C.base = type { i32 (...)** }
+
+// Class A contains a vtable ptr, then int, then padding
+// CHECK: %class.A = type <{ i32 (...)**, i32, [4 x i8] }>
+
+// VTable for B. Contains an extra field at the start for the virtual-base offset.
+// CHECK: @_ZTV1B.local = private unnamed_addr constant { [4 x i32], [4 x i32] } { [4 x i32] [i32 8, i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64 }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B4barBEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 3) to i64)) to i32)], [4 x i32] [i32 0, i32 -8, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64 }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1B.local, i32 0, i32 1, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1B.local, i32 0, i32 1, i32 3) to i64)) to i32)] }, align 4
+
+// VTT for B
+// CHECK: @_ZTT1B = unnamed_addr constant [2 x i8*] [i8* bitcast (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1B.local, i32 0, inrange i32 0, i32 3) to i8*), i8* bitcast (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1B.local, i32 0, inrange i32 1, i32 3) to i8*)], align 8
+
+// VTable for C
+// CHECK: @_ZTV1C.local = private unnamed_addr constant { [4 x i32], [4 x i32] } { [4 x i32] [i32 8, i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64 }** @_ZTI1C.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1C.local, i32 0, i32 0, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.C*)* @_ZN1C4barCEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1C.local, i32 0, i32 0, i32 3) to i64)) to i32)], [4 x i32] [i32 0, i32 -8, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64 }** @_ZTI1C.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1C.local, i32 0, i32 1, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1C.local, i32 0, i32 1, i32 3) to i64)) to i32)] }, align 4
+
+// VTT for C
+// CHECK: @_ZTT1C = unnamed_addr constant [2 x i8*] [i8* bitcast (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1C.local, i32 0, inrange i32 0, i32 3) to i8*), i8* bitcast (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1C.local, i32 0, inrange i32 1, i32 3) to i8*)], align 8
+
+// VTable for D
+// CHECK: @_ZTV1D.local = private unnamed_addr constant { [5 x i32], [4 x i32], [4 x i32] } { [5 x i32] [i32 16, i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64, i8*, i64 }** @_ZTI1D.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 0, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B4barBEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 0, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.D*)* @_ZN1D3bazEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 0, i32 3) to i64)) to i32)], [4 x i32] [i32 8, i32 -8, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64, i8*, i64 }** @_ZTI1D.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 1, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.C*)* @_ZN1C4barCEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 1, i32 3) to i64)) to i32)], [4 x i32] [i32 0, i32 -16, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64, i8*, i64 }** @_ZTI1D.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 2, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, i32 2, i32 3) to i64)) to i32)] }, align 4
+
+// VTT for D
+// CHECK: @_ZTT1D = unnamed_addr constant [7 x i8*] [i8* bitcast (i32* getelementptr inbounds ({ [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, inrange i32 0, i32 3) to i8*), i8* bitcast (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D0_1B.local, i32 0, inrange i32 0, i32 3) to i8*), i8* bitcast (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D0_1B.local, i32 0, inrange i32 1, i32 3) to i8*), i8* bitcast (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D8_1C.local, i32 0, inrange i32 0, i32 3) to i8*), i8* bitcast (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D8_1C.local, i32 0, inrange i32 1, i32 3) to i8*), i8* bitcast (i32* getelementptr inbounds ({ [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, inrange i32 2, i32 3) to i8*), i8* bitcast (i32* getelementptr inbounds ({ [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local, i32 0, inrange i32 1, i32 3) to i8*)], align 8
+
+// Construction vtable for B-in-D
+// CHECK: @_ZTC1D0_1B.local = private unnamed_addr constant { [4 x i32], [4 x i32] } { [4 x i32] [i32 16, i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64 }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D0_1B.local, i32 0, i32 0, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B4barBEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D0_1B.local, i32 0, i32 0, i32 3) to i64)) to i32)], [4 x i32] [i32 0, i32 -16, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64 }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D0_1B.local, i32 0, i32 1, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D0_1B.local, i32 0, i32 1, i32 3) to i64)) to i32)] }, align 4
+
+// Construction vtable for C-in-D
+// CHECK: @_ZTC1D8_1C.local = private unnamed_addr constant { [4 x i32], [4 x i32] } { [4 x i32] [i32 8, i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64 }** @_ZTI1C.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D8_1C.local, i32 0, i32 0, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.C*)* @_ZN1C4barCEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D8_1C.local, i32 0, i32 0, i32 3) to i64)) to i32)], [4 x i32] [i32 0, i32 -8, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64 }** @_ZTI1C.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D8_1C.local, i32 0, i32 1, i32 3) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D8_1C.local, i32 0, i32 1, i32 3) to i64)) to i32)] }, align 4
+
+// CHECK: @_ZTV1B = unnamed_addr alias { [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1B.local
+// CHECK: @_ZTV1C = unnamed_addr alias { [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTV1C.local
+// CHECK: @_ZTC1D0_1B = unnamed_addr alias { [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D0_1B.local
+// CHECK: @_ZTC1D8_1C = unnamed_addr alias { [4 x i32], [4 x i32] }, { [4 x i32], [4 x i32] }* @_ZTC1D8_1C.local
+// CHECK: @_ZTV1D = unnamed_addr alias { [5 x i32], [4 x i32], [4 x i32] }, { [5 x i32], [4 x i32], [4 x i32] }* @_ZTV1D.local
+
+// CHECK:      define void @_Z5D_fooP1D(%class.D* %d) local_unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   [[d:%[0-9]+]] = bitcast %class.D* %d to i8**
+// CHECK-NEXT:   [[vtable:%[a-z0-9]+]] = load i8*, i8** [[d]], align 8
+
+// This normally would've been -24 (8 bytes for the virtual call/base offset, 8
+// bytes for the offset to top, and 8 bytes for the RTTI pointer), but since all
+// components in the vtable are halved to 4 bytes, the offset is halved to -12.
+// CHECK-NEXT:   [[vbase_offset_ptr:%[a-z0-9.]+]] = getelementptr i8, i8* [[vtable]], i64 -12
+
+// CHECK-NEXT:   [[vbase_offset_ptr2:%[a-z0-9.]+]] = bitcast i8* [[vbase_offset_ptr]] to i32*
+// CHECK-NEXT:   [[vbase_offset:%[a-z0-9.]+]] = load i32, i32* [[vbase_offset_ptr2]], align 4
+// CHECK-NEXT:   [[d:%[0-9]+]] = bitcast %class.D* %d to i8*
+// CHECK-NEXT:   [[vbase_offset2:%.+]] = sext i32 [[vbase_offset]] to i64
+// CHECK-NEXT:   [[add_ptr:%[a-z0-9.]+]] = getelementptr inbounds i8, i8* [[d]], i64 [[vbase_offset2]]
+// CHECK-NEXT:   [[a:%[0-9]+]] = bitcast i8* [[add_ptr]] to %class.A*
+// CHECK-NEXT:   [[a_i8_ptr:%[0-9]+]] = bitcast i8* [[add_ptr]] to i8**
+// CHECK-NEXT:   [[vtable:%[a-z0-9]+]] = load i8*, i8** [[a_i8_ptr]], align 8
+// CHECK-NEXT:   [[ptr:%[0-9]+]] = call i8* @llvm.load.relative.i32(i8* [[vtable]], i32 0)
+// CHECK-NEXT:   [[method:%[0-9]+]] = bitcast i8* [[ptr]] to void (%class.A*)*
+// CHECK-NEXT:   call void [[method]](%class.A* [[a]])
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  virtual void foo();
+  int a;
+};
+
+class B : public virtual A {
+public:
+  virtual void barB();
+};
+
+class C : public virtual A {
+  virtual void barC();
+};
+
+class D : public B, C {
+public:
+  virtual void baz();
+};
+
+void B::barB() {}
+void C::barC() {}
+void D::baz() {}
+
+void D_foo(D *d) {
+  d->foo();
+}

clang/test/CodeGenCXX/Fuchsia/dynamic-cast.cpp

+// dynamic_cast
+// Ensure that dynamic casting works normally
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O3 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK:      define %class.A* @_Z6upcastP1B(%class.B* readnone %b) local_unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   [[a:%[0-9]+]] = getelementptr %class.B, %class.B* %b, i64 0, i32 0
+// CHECK-NEXT:   ret %class.A* [[a]]
+// CHECK-NEXT: }
+
+// CHECK:      define %class.B* @_Z8downcastP1A(%class.A* readonly %a) local_unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   [[isnull:%[0-9]+]] = icmp eq %class.A* %a, null
+// CHECK-NEXT:   br i1 [[isnull]], label %[[dynamic_cast_end:[a-z0-9._]+]], label %[[dynamic_cast_notnull:[a-z0-9._]+]]
+// CHECK:      [[dynamic_cast_notnull]]:
+// CHECK-NEXT:   [[a:%[0-9]+]] = bitcast %class.A* %a to i8*
+// CHECK-NEXT:   [[as_b:%[0-9]+]] = tail call i8* @__dynamic_cast(i8* nonnull [[a]], i8* bitcast ({ i8*, i8* }* @_ZTI1A to i8*), i8* bitcast ({ i8*, i8*, i8* }* @_ZTI1B to i8*), i64 0)
+// CHECK-NEXT:   [[b:%[0-9]+]] = bitcast i8* [[as_b]] to %class.B*
+// CHECK-NEXT:   br label %[[dynamic_cast_end]]
+// CHECK:      [[dynamic_cast_end]]:
+// CHECK-NEXT:   [[res:%[0-9]+]] = phi %class.B* [ [[b]], %[[dynamic_cast_notnull]] ], [ null, %entry ]
+// CHECK-NEXT:   ret %class.B* [[res]]
+// CHECK-NEXT: }
+
+// CHECK: declare i8* @__dynamic_cast(i8*, i8*, i8*, i64) local_unnamed_addr
+
+// CHECK:      define %class.B* @_Z8selfcastP1B(%class.B* readnone returned %b) local_unnamed_addr
+// CHECK-NEXT: entry
+// CHECK-NEXT:   ret %class.B* %b
+// CHECK-NEXT: }
+
+// CHECK: define i8* @_Z9void_castP1B(%class.B* readonly %b) local_unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   [[isnull:%[0-9]+]] = icmp eq %class.B* %b, null
+// CHECK-NEXT:   br i1 [[isnull]], label %[[dynamic_cast_end:[a-z0-9._]+]], label %[[dynamic_cast_notnull:[a-z0-9._]+]]
+// CHECK:      [[dynamic_cast_notnull]]:
+// CHECK-NEXT:   [[b2:%[0-9]+]] = bitcast %class.B* %b to i32**
+// CHECK-NEXT:   [[vtable:%[a-z0-9]+]] = load i32*, i32** [[b2]], align 8
+// CHECK-NEXT:   [[offset_ptr:%.+]] = getelementptr inbounds i32, i32* [[vtable]], i64 -2
+// CHECK-NEXT:   [[offset_to_top:%.+]] = load i32, i32* [[offset_ptr]], align 4
+// CHECK-NEXT:   [[b:%[0-9]+]] = bitcast %class.B* %b to i8*
+// CHECK-NEXT:   [[offset_to_top2:%.+]] = sext i32 [[offset_to_top]] to i64
+// CHECK-NEXT:   [[casted:%.+]] = getelementptr inbounds i8, i8* [[b]], i64 [[offset_to_top2]]
+// CHECK-NEXT:   br label %[[dynamic_cast_end]]
+// CHECK:      [[dynamic_cast_end]]:
+// CHECK-NEXT:   [[res:%[0-9]+]] = phi i8* [ [[casted]], %[[dynamic_cast_notnull]] ], [ null, %entry ]
+// CHECK-NEXT:   ret i8* [[res]]
+// CHECK-NEXT: }
+
+class A {
+public:
+  virtual void foo();
+};
+
+class B : public A {
+public:
+  void foo() override;
+};
+
+void A::foo() {}
+void B::foo() {}
+
+A *upcast(B *b) {
+  return dynamic_cast<A *>(b);
+}
+
+B *downcast(A *a) {
+  return dynamic_cast<B *>(a);
+}
+
+B *selfcast(B *b) {
+  return dynamic_cast<B *>(b);
+}
+
+void *void_cast(B *b) {
+  return dynamic_cast<void *>(b);
+}

clang/test/CodeGenCXX/Fuchsia/inheritted-virtual-function.cpp

+// Check the layout of the vtable for a child class that inherits a virtual
+// function but does not override it.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+class A {
+public:
+  virtual void foo();
+};
+
+// The VTable for B should look similar to the vtable for A but the component for foo() should point to A::foo() and the component for bar() should point to B::bar().
+// CHECK: @_ZTV1B.local = private unnamed_addr constant { [4 x i32] } { [4 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i8* }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B3barEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+// CHECK: @_ZTV1B = unnamed_addr alias { [4 x i32] }, { [4 x i32] }* @_ZTV1B.local
+
+class B : public A {
+public:
+  virtual void bar();
+};
+
+void A::foo() {}
+void B::bar() {}
+
+void A_foo(A *a) {
+  a->foo();
+}
+
+void B_foo(B *b) {
+  b->foo();
+}

clang/test/CodeGenCXX/Fuchsia/inline-virtual-function.cpp

+// The VTable is not in a comdat but the inline methods are.
+// This doesn’t affect the vtable or the stubs we emit.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: $_ZN1A3fooEv.stub = comdat any
+// CHECK: $_ZN1A3barEv.stub = comdat any
+// CHECK: $_ZTI1A.rtti_proxy = comdat any
+
+// The vtable has a key function (A::foo()) so it does not have a comdat
+// CHECK: @_ZTV1A.local = private unnamed_addr constant { [4 x i32] } { [4 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8* }** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3barEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+// CHECK: @_ZTV1A = unnamed_addr alias { [4 x i32] }, { [4 x i32] }* @_ZTV1A.local
+
+// We do not declare the stub with linkonce_odr so it can be emitted as .globl.
+// CHECK:      define hidden void @_ZN1A3barEv.stub(%class.A* %0) unnamed_addr #{{[0-9]+}} comdat {
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   tail call void @_ZN1A3barEv(%class.A* %0)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  virtual void foo(); // Key func
+  inline virtual void bar() {}
+};
+
+void A::foo() {}

clang/test/CodeGenCXX/Fuchsia/inlined-key-function.cpp

+// Inline comdat method definition example.
+// The VTable is in a comdat and defined anywhere the inline definition is.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: $_ZN1A3fooEv.stub = comdat any
+// CHECK: $_ZTV1A = comdat any
+// CHECK: $_ZTS1A = comdat any
+// CHECK: $_ZTI1A = comdat any
+// CHECK: $_ZTI1A.rtti_proxy = comdat any
+
+// The VTable is linkonce_odr and in a comdat here bc it’s key function is inline defined.
+// CHECK: @_ZTV1A.local = linkonce_odr hidden unnamed_addr constant { [3 x i32] } { [3 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8* }** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, comdat($_ZTV1A), align 4
+// CHECK: @_ZTV1A = linkonce_odr unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1A.local
+
+// CHECK:      define hidden void @_ZN1A3fooEv.stub(%class.A* %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   tail call void @_ZN1A3fooEv(%class.A* %0)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  virtual void foo();
+};
+void A_foo(A *a);
+
+inline void A::foo() {}
+
+void func() {
+  A a;
+  A_foo(&a);
+}

clang/test/CodeGenCXX/Fuchsia/member-function-pointer.cpp

+// Member pointer to virtual function.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O3 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK:      define void @_Z4funcP1AMS_FvvE(%class.A* %a, [2 x i64] %fn.coerce) local_unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   [[fn_ptr:%.+]] = extractvalue [2 x i64] %fn.coerce, 0
+// CHECK-NEXT:   [[adjust:%.+]] = extractvalue [2 x i64] %fn.coerce, 1
+// CHECK-NEXT:   [[this:%.+]] = bitcast %class.A* %a to i8*
+// CHECK-NEXT:   [[this_adj:%.+]] = getelementptr inbounds i8, i8* [[this]], i64 [[adjust]]
+// CHECK-NEXT:   [[a:%.+]] = bitcast i8* [[this_adj]] to %class.A*
+// CHECK-NEXT:   [[virtbit:%.+]] = and i64 [[fn_ptr]], 1
+// CHECK-NEXT:   [[isvirt:%.+]] = icmp eq i64 [[virtbit]], 0
+// CHECK-NEXT:   br i1 [[isvirt]], label %[[nonvirt:.+]], label %[[virt:.+]]
+// CHECK:      [[virt]]:
+
+// The loading of the virtual function here should be replaced with a llvm.load.relative() call.
+// CHECK-NEXT:   [[this:%.+]] = bitcast i8* [[this_adj]] to i8**
+// CHECK-NEXT:   [[vtable:%.+]] = load i8*, i8** [[this]], align 8
+// CHECK-NEXT:   [[offset:%.+]] = add i64 [[fn_ptr]], -1
+// CHECK-NEXT:   [[ptr:%.+]] = tail call i8* @llvm.load.relative.i64(i8* [[vtable]], i64 [[offset]])
+// CHECK-NEXT:   [[method:%.+]] = bitcast i8* [[ptr]] to void (%class.A*)*
+// CHECK-NEXT:   br label %[[memptr_end:.+]]
+// CHECK:      [[nonvirt]]:
+// CHECK-NEXT:   [[method2:%.+]] = inttoptr i64 [[fn_ptr]] to void (%class.A*)*
+// CHECK-NEXT:   br label %[[memptr_end]]
+// CHECK:      [[memptr_end]]:
+// CHECK-NEXT:   [[method3:%.+]] = phi void (%class.A*)* [ [[method]], %[[virt]] ], [ [[method2]], %[[nonvirt]] ]
+// CHECK-NEXT:   tail call void [[method3]](%class.A* [[a]])
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  virtual void foo();
+};
+
+class B : public A {
+public:
+  void foo() override;
+};
+
+typedef void (A::*A_foo)();
+
+void func(A *a, A_foo fn) {
+  (a->*fn)();
+}

clang/test/CodeGenCXX/Fuchsia/multiple-inheritance.cpp

+// Multiple inheritance.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: %class.C = type { %class.A, %class.B }
+// CHECK: %class.A = type { i32 (...)** }
+// CHECK: %class.B = type { i32 (...)** }
+
+// VTable for C contains 2 sub-vtables (represented as 2 structs). The first contains the components for B and the second contains the components for C. The RTTI ptr in both arrays still point to the RTTI struct for C.
+// The component for bar() instead points to a thunk which redirects to C::bar() which overrides B::bar().
+// Now that we have a class with 2 parents, the offset to top in the second array is non-zero.
+// CHECK: @_ZTV1C.local = private unnamed_addr constant { [4 x i32], [3 x i32] } { [4 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64, i8*, i64 }** @_ZTI1C.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [3 x i32] }, { [4 x i32], [3 x i32] }* @_ZTV1C.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.C*)* @_ZN1C3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [3 x i32] }, { [4 x i32], [3 x i32] }* @_ZTV1C.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.C*)* @_ZN1C3barEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [3 x i32] }, { [4 x i32], [3 x i32] }* @_ZTV1C.local, i32 0, i32 0, i32 2) to i64)) to i32)], [3 x i32] [i32 -8, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i32, i32, i8*, i64, i8*, i64 }** @_ZTI1C.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [3 x i32] }, { [4 x i32], [3 x i32] }* @_ZTV1C.local, i32 0, i32 1, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.C*)* @_ZThn8_N1C3barEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32], [3 x i32] }, { [4 x i32], [3 x i32] }* @_ZTV1C.local, i32 0, i32 1, i32 2) to i64)) to i32)] }, align 4
+
+// CHECK: @_ZTV1C = unnamed_addr alias { [4 x i32], [3 x i32] }, { [4 x i32], [3 x i32] }* @_ZTV1C.local
+
+// CHECK:      define void @_Z8C_foobarP1C(%class.C* %c) local_unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   [[c:%[0-9]+]] = bitcast %class.C* %c to i8**
+// CHECK-NEXT:   [[vtable:%[a-z0-9]+]] = load i8*, i8** [[c]], align 8
+
+// Offset 0 to get first method
+// CHECK-NEXT:   [[ptr1:%[0-9]+]] = call i8* @llvm.load.relative.i32(i8* [[vtable]], i32 0)
+// CHECK-NEXT:   [[method1:%[0-9]+]] = bitcast i8* [[ptr1]] to void (%class.C*)*
+// CHECK-NEXT:   call void [[method1]](%class.C* %c)
+// CHECK-NEXT:   [[vtable:%[a-z0-9]+]] = load i8*, i8** [[c]], align 8
+
+// Offset by 4 to get the next bar()
+// CHECK-NEXT:   [[ptr2:%[0-9]+]] = call i8* @llvm.load.relative.i32(i8* [[vtable]], i32 4)
+// CHECK-NEXT:   [[method2:%[0-9]+]] = bitcast i8* [[ptr2]] to void (%class.C*)*
+// CHECK-NEXT:   call void [[method2]](%class.C* %c)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  virtual void foo();
+};
+
+class B {
+  virtual void bar();
+};
+
+class C : public A, public B {
+public:
+  void foo() override;
+  void bar() override;
+};
+
+void C::foo() {}
+void C::bar() {}
+
+void C_foobar(C *c) {
+  c->foo();
+  c->bar();
+}

clang/test/CodeGenCXX/Fuchsia/no-alias-when-dso-local.cpp

+// Check that no alias is emitted when the vtable is already dso_local. This can
+// happen if the class is hidden.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s --check-prefix=DEFAULT-VIS
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables -fvisibility hidden | FileCheck %s --check-prefix=HIDDEN-VIS
+
+// DEFAULT-VIS: @_ZTV1A.local = private unnamed_addr constant
+// DEFAULT-VIS: @_ZTV1A = unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1A.local
+// HIDDEN-VIS-NOT: @_ZTV1A.local
+// HIDDEN-VIS: @_ZTV1A = hidden unnamed_addr constant
+class A {
+ public:
+  virtual void func();
+};
+
+void A::func() {}

clang/test/CodeGenCXX/Fuchsia/no-stub-when-dso-local.cpp

+// Check that we do not emit a stub for a virtual function if it is already
+// known to be in the same linkage unit as the vtable.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s --check-prefixes=CHECK,NO-OPT
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables -O3 | FileCheck %s --check-prefixes=CHECK,OPT
+
+// The vtable offset is relative to _ZN1A3fooEv instead of a stub. We can do
+// this since hidden functions are implicitly dso_local.
+// CHECK: @_ZTV1A.local = private unnamed_addr constant { [5 x i32] } { [5 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8* }** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32] }, { [5 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32] }, { [5 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (%class.A* (%class.A*)* @_ZN1AD1Ev.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32] }, { [5 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1AD0Ev.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32] }, { [5 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+
+// Despite the complete object destructor being hidden, we should still emit a
+// stub for it because it's possible, when optimizations are enabled, for the
+// dtor to be "devirtualized" into the destructor for a parent class if the
+// child class doesn't implement its own dtor. For complete destructors, we
+// always emit and use a stub.
+// @_ZTV1B = hidden unnamed_addr constant { [5 x i32] } { [5 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i8* }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32] }, { [5 x i32] }* @_ZTV1B, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B3fooEv to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32] }, { [5 x i32] }* @_ZTV1B, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (%class.B* (%class.B*)* @_ZN1BD1Ev.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32] }, { [5 x i32] }* @_ZTV1B, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1BD0Ev to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [5 x i32] }, { [5 x i32] }* @_ZTV1B, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+
+// CHECK: @_ZTV1A = unnamed_addr alias { [5 x i32] }, { [5 x i32] }* @_ZTV1A.local
+
+// CHECK: @_ZN1A3fooEv
+// CHECK-NOT: @_ZN1A3fooEv.stub
+
+// The complete object destructor is hidden.
+// NO-OPT: define linkonce_odr hidden %class.B* @_ZN1BD1Ev
+// OPT-NOT: @_ZN1BD1Ev
+// CHECK: @_ZN1BD1Ev.stub
+
+#define HIDDEN __attribute__((visibility("hidden")))
+
+class A {
+public:
+  HIDDEN virtual void foo();
+  virtual ~A();
+};
+
+class HIDDEN B : public A {
+public:
+  virtual void foo();
+};
+
+void A::foo() {}
+
+void A_foo(A *a) {
+  a->foo();
+}
+
+void B::foo() {
+  A::foo();
+}

clang/test/CodeGenCXX/Fuchsia/override-pure-virtual-method.cpp

+// Override pure virtual function.
+// We instead emit zero for the pure virtual function component. See PR43094 for
+// details.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: @_ZTV1A.local = private unnamed_addr constant { [4 x i32] } { [4 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8* }** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 0, i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3barEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+
+// CHECK: @_ZTV1B.local = private unnamed_addr constant { [4 x i32] } { [4 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i8* }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B3barEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+
+// CHECK: @_ZTV1A = unnamed_addr alias { [4 x i32] }, { [4 x i32] }* @_ZTV1A.local
+// CHECK: @_ZTV1B = unnamed_addr alias { [4 x i32] }, { [4 x i32] }* @_ZTV1B.local
+
+// CHECK-NOT: declare void @__cxa_pure_virtual() unnamed_addr
+
+class A {
+public:
+  virtual void foo() = 0;
+  virtual void bar();
+};
+
+class B : public A {
+public:
+  void foo() override;
+  void bar() override;
+};
+
+void A::bar() {}
+void B::foo() {}
+void B::bar() {}
+
+void A_foo(A *a) {
+  a->foo();
+}

clang/test/CodeGenCXX/Fuchsia/overriden-virtual-function.cpp

+// Check the layout of the vtable for a child class that inherits a virtual
+// function but does override it.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: @_ZTV1B.local = private unnamed_addr constant { [4 x i32] } { [4 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i8* }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B3barEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [4 x i32] }, { [4 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+
+// CHECK: @_ZTV1B = unnamed_addr alias { [4 x i32] }, { [4 x i32] }* @_ZTV1B.local
+
+class A {
+public:
+  virtual void foo();
+};
+
+class B : public A {
+public:
+  void foo() override;
+  virtual void bar();
+};
+
+void A::foo() {}
+void B::foo() {}
+
+void A_foo(A *a) {
+  a->foo();
+}
+
+void B_foo(B *b) {
+  b->foo();
+}

clang/test/CodeGenCXX/Fuchsia/parent-and-child-in-comdats.cpp

+// Cross comdat example
+// Both the parent and child VTablea are in their own comdat sections.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// Comdats are emitted for both A and B in this module and for their respective implementations of foo().
+// CHECK: $_ZN1A3fooEv = comdat any
+// CHECK: $_ZN1A3fooEv.stub = comdat any
+// CHECK: $_ZN1B3fooEv = comdat any
+// CHECK: $_ZN1B3fooEv.stub = comdat any
+// CHECK: $_ZTV1A = comdat any
+// CHECK: $_ZTS1A = comdat any
+// CHECK: $_ZTI1A = comdat any
+// CHECK: $_ZTI1A.rtti_proxy = comdat any
+// CHECK: $_ZTV1B = comdat any
+// CHECK: $_ZTS1B = comdat any
+// CHECK: $_ZTI1B = comdat any
+// CHECK: $_ZTI1B.rtti_proxy = comdat any
+
+// Both the vtables for A and B are emitted and in their own comdats.
+// CHECK: @_ZTV1A.local = linkonce_odr hidden unnamed_addr constant { [3 x i32] } { [3 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8* }** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, comdat($_ZTV1A), align 4
+// CHECK: @_ZTV1B.local = linkonce_odr hidden unnamed_addr constant { [3 x i32] } { [3 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8*, i8* }** @_ZTI1B.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.B*)* @_ZN1B3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1B.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, comdat($_ZTV1B), align 4
+
+// CHECK: @_ZTV1A = linkonce_odr unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1A.local
+// CHECK: @_ZTV1B = linkonce_odr unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1B.local
+
+// CHECK: declare void @_Z5A_fooP1A(%class.A*)
+
+// The stubs and implementations for foo() are in their own comdat sections.
+// CHECK:      define linkonce_odr void @_ZN1A3fooEv(%class.A* %this) unnamed_addr #{{[0-9]+}} comdat
+
+// CHECK:      define hidden void @_ZN1A3fooEv.stub(%class.A* %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   tail call void @_ZN1A3fooEv(%class.A* %0)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+// CHECK:      define linkonce_odr void @_ZN1B3fooEv(%class.B* %this) unnamed_addr #{{[0-9]+}} comdat
+
+// CHECK:      define hidden void @_ZN1B3fooEv.stub(%class.B* %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   tail call void @_ZN1B3fooEv(%class.B* %0)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  inline virtual void foo() {}
+};
+class B : public A {
+public:
+  inline void foo() override {}
+};
+void A_foo(A *a);
+
+// func() is used so that the vtable for B is accessed when creating the instance.
+void func() {
+  A a;
+  B b;
+  A_foo(&a);
+  A_foo(&b);
+}

clang/test/CodeGenCXX/Fuchsia/parent-vtable-in-comdat.cpp

+// Cross comdat example
+// Parent VTable is in a comdat section.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// A::foo() has a comdat since it is an inline function
+// CHECK: $_ZN1A3fooEv = comdat any
+// CHECK: $_ZN1A3fooEv.stub = comdat any
+// CHECK: $_ZTV1A = comdat any
+// CHECK: $_ZTS1A = comdat any
+
+// The VTable for A has its own comdat section bc it has no key function
+// CHECK: $_ZTI1A = comdat any
+// CHECK: $_ZTI1A.rtti_proxy = comdat any
+
+// The VTable for A is emitted here and in a comdat section since it has no key function, and is used in this module when creating an instance of A.
+// CHECK: @_ZTV1A.local = linkonce_odr hidden unnamed_addr constant { [3 x i32] } { [3 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8* }** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, comdat($_ZTV1A), align 4
+// CHECK: @_ZTVN10__cxxabiv117__class_type_infoE = external global i8*
+// CHECK: @_ZTS1A = linkonce_odr constant [3 x i8] c"1A\00", comdat, align 1
+// CHECK: @_ZTI1A = linkonce_odr constant { i8*, i8* } { i8* getelementptr inbounds (i8, i8* bitcast (i8** @_ZTVN10__cxxabiv117__class_type_infoE to i8*), i32 8), i8* getelementptr inbounds ([3 x i8], [3 x i8]* @_ZTS1A, i32 0, i32 0) }, comdat, align 8
+// CHECK: @_ZTI1A.rtti_proxy = hidden unnamed_addr constant { i8*, i8* }* @_ZTI1A, comdat
+// CHECK: @_ZTV1A = linkonce_odr unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1A.local
+
+// CHECK:      define linkonce_odr void @_ZN1A3fooEv(%class.A* %this) unnamed_addr #{{[0-9]+}} comdat
+
+// CHECK:      define hidden void @_ZN1A3fooEv.stub(%class.A* %0) unnamed_addr #{{[0-9]+}} comdat {
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   tail call void @_ZN1A3fooEv(%class.A* %0)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  inline virtual void foo() {}
+};
+class B : public A {
+public:
+  void foo() override;
+};
+void A_foo(A *a);
+
+void A_foo(A *a) { a->foo(); }
+
+// func() is only used to emit a vtable.
+void func() {
+  A a;
+  A_foo(&a);
+}

clang/test/CodeGenCXX/Fuchsia/pass-byval-attributes.cpp

+// ByVal attributes should propogate through to produce proper assembly and
+// avoid "unpacking" structs within the stubs on x86_64.
+
+// RUN: %clang_cc1 %s -triple=x86_64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+struct LargeStruct {
+  char x[24];
+  virtual ~LargeStruct() {}
+};
+
+struct fidl_string {
+  unsigned long long size;
+  char *data;
+};
+static_assert(sizeof(fidl_string) == 16, "");
+
+class Base {
+public:
+  virtual void func(LargeStruct, fidl_string, LargeStruct, fidl_string) = 0;
+};
+
+class Derived : public Base {
+public:
+  void func(LargeStruct, fidl_string, LargeStruct, fidl_string) override;
+};
+
+// The original function takes a byval pointer.
+// CHECK: define void @_ZN7Derived4funcE11LargeStruct11fidl_stringS0_S1_(%class.Derived* %this, %struct.LargeStruct* %ls, i64 %sv1.coerce0, i8* %sv1.coerce1, %struct.LargeStruct* %ls2, %struct.fidl_string* byval(%struct.fidl_string) align 8 %sv2) unnamed_addr
+
+// So the stub should take and pass one also.
+// CHECK:      define hidden void @_ZN7Derived4funcE11LargeStruct11fidl_stringS0_S1_.stub(%class.Derived* %0, %struct.LargeStruct* %1, i64 %2, i8* %3, %struct.LargeStruct* %4, %struct.fidl_string* byval(%struct.fidl_string) align 8 %5) unnamed_addr {{#[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   tail call void @_ZN7Derived4funcE11LargeStruct11fidl_stringS0_S1_(%class.Derived* %0, %struct.LargeStruct* %1, i64 %2, i8* %3, %struct.LargeStruct* %4, %struct.fidl_string* byval(%struct.fidl_string) align 8 %5)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+void Derived::func(LargeStruct ls, fidl_string sv1, LargeStruct ls2, fidl_string sv2) {}

clang/test/CodeGenCXX/Fuchsia/relative-vtables-flag.cpp

+// Check the layout of the vtable for a normal class.
+// The Fuchsia relative vtables ABI will be hidden behind a flag for now as part
+// of a soft incremental rollout. This ABI should only be used if the flag for
+// it is passed on Fuchsia.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck --check-prefix=RELATIVE-ABI %s
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm -fno-experimental-relative-c++-abi-vtables | FileCheck --check-prefix=DEFAULT-ABI %s
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm | FileCheck --check-prefix=DEFAULT-ABI %s
+
+// VTable contains offsets and references to the hidden symbols
+// RELATIVE-ABI: @_ZTV1A.local = private unnamed_addr constant { [3 x i32] } { [3 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8* }** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+// RELATIVE-ABI: @_ZTV1A = unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1A.local
+// DEFAULT-ABI: @_ZTV1A = unnamed_addr constant { [3 x i8*] } { [3 x i8*] [i8* null, i8* bitcast ({ i8*, i8* }* @_ZTI1A to i8*), i8* bitcast (void (%class.A*)* @_ZN1A3fooEv to i8*)] }, align 8
+
+class A {
+public:
+  virtual void foo();
+};
+
+void A::foo() {}
+
+void A_foo(A *a) {
+  a->foo();
+}

clang/test/CodeGenCXX/Fuchsia/simple-vtable-definition.cpp

+// Check the layout of the vtable for a normal class.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// We should be emitting comdats for each of the virtual function stubs and RTTI proxies
+// CHECK: $_ZN1A3fooEv.stub = comdat any
+// CHECK: $_ZTI1A.rtti_proxy = comdat any
+
+// VTable contains offsets and references to the hidden symbols
+// The vtable definition itself is private so we can take relative references to
+// it. The vtable symbol will be exposed through a public alias.
+// CHECK: @_ZTV1A.local = private unnamed_addr constant { [3 x i32] } { [3 x i32] [i32 0, i32 trunc (i64 sub (i64 ptrtoint ({ i8*, i8* }** @_ZTI1A.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32), i32 trunc (i64 sub (i64 ptrtoint (void (%class.A*)* @_ZN1A3fooEv.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV1A.local, i32 0, i32 0, i32 2) to i64)) to i32)] }, align 4
+// CHECK: @_ZTVN10__cxxabiv117__class_type_infoE = external global i8*
+// CHECK: @_ZTS1A = constant [3 x i8] c"1A\00", align 1
+// CHECK: @_ZTI1A = constant { i8*, i8* } { i8* getelementptr inbounds (i8, i8* bitcast (i8** @_ZTVN10__cxxabiv117__class_type_infoE to i8*), i32 8), i8* getelementptr inbounds ([3 x i8], [3 x i8]* @_ZTS1A, i32 0, i32 0) }, align 8
+
+// The stub should be in a comdat
+// CHECK: @_ZTI1A.rtti_proxy = hidden unnamed_addr constant { i8*, i8* }* @_ZTI1A, comdat
+
+// The vtable symbol is exposed through an alias.
+// @_ZTV1A = dso_local unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1A.local
+
+// CHECK:      define void @_ZN1A3fooEv(%class.A* nocapture %this) unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+// This function should be in a comdat
+// CHECK:      define hidden void @_ZN1A3fooEv.stub(%class.A* nocapture %0) unnamed_addr #{{[0-9]+}} comdat
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  virtual void foo();
+};
+
+void A::foo() {}
+
+void A_foo(A *a) {
+  a->foo();
+}

clang/test/CodeGenCXX/Fuchsia/stub-linkages.cpp

+// If the linkage of the class is internal, then the stubs and proxies should
+// also be internally linked.
+
+// RUN: %clang_cc1 %s -triple=x86_64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// External linkage.
+// CHECK: @_ZTI8External.rtti_proxy = hidden unnamed_addr constant { i8*, i8* }* @_ZTI8External, comdat
+
+class External {
+public:
+  virtual void func();
+};
+
+void External::func() {}
+
+// Internal linkage.
+// CHECK: @_ZTIN12_GLOBAL__N_18InternalE.rtti_proxy = internal unnamed_addr constant { i8*, i8* }* @_ZTIN12_GLOBAL__N_18InternalE
+namespace {
+
+class Internal {
+public:
+  virtual void func();
+};
+
+void Internal::func() {}
+
+} // namespace
+
+// This gets the same treatment as an externally available vtable.
+// CHECK: @_ZTI11LinkOnceODR.rtti_proxy = hidden unnamed_addr constant { i8*, i8* }* @_ZTI11LinkOnceODR, comdat
+class LinkOnceODR {
+public:
+  virtual void func() {} // A method defined in the class definition results in this linkage for the vtable.
+};
+
+// Force an emission of a vtable for Internal by using it here.
+void manifest_internal() {
+  Internal internal;
+  (void)internal;
+  LinkOnceODR linkonceodr;
+  (void)linkonceodr;
+}
+
+// Aliases are typically emitted after the vtable definitions but before the
+// function definitions.
+// CHECK: @_ZTV8External = unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV8External.local
+// CHECK: @_ZTV11LinkOnceODR = linkonce_odr unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV11LinkOnceODR.local
+
+// CHECK: define void @_ZN8External4funcEv
+// CHECK: define internal void @_ZN12_GLOBAL__N_18Internal4funcEv.stub
+// CHECK: define hidden void @_ZN11LinkOnceODR4funcEv.stub

clang/test/CodeGenCXX/Fuchsia/thunk-mangling.cpp

+// Check that virtual thunks are unaffected by the relative ABI.
+// The offset of thunks is mangled into the symbol name, which could result in
+// linking errors for binaries that want to look for symbols in SOs made with
+// this ABI.
+// Running that linked binary still won't work since we're using conflicting
+// ABIs, but we should still be able to link.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O1 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// This would be normally n24 (3 ptr widths) but is 12 since the vtable is
+// entierely made of i32s now.
+// CHECK: _ZTv0_n12_N7Derived1fEi
+
+class Base {
+public:
+  virtual int f(int x);
+
+private:
+  long x;
+};
+
+class Derived : public virtual Base {
+public:
+  virtual int f(int x);
+
+private:
+  long y;
+};
+
+int Base::f(int x) { return x + 1; }
+int Derived::f(int x) { return x + 2; }

clang/test/CodeGenCXX/Fuchsia/type-info.cpp

+// Check typeid() + type_info
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O3 -S -o - -emit-llvm -fcxx-exceptions -fexceptions -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: %class.A = type { i32 (...)** }
+// CHECK: %class.B = type { %class.A }
+// CHECK: %"class.std::type_info" = type { i32 (...)**, i8* }
+
+// CHECK: $_ZN1A3fooEv.stub = comdat any
+// CHECK: $_ZN1B3fooEv.stub = comdat any
+// CHECK: $_ZTI1A.rtti_proxy = comdat any
+// CHECK: $_ZTI1B.rtti_proxy = comdat any
+
+// CHECK: @_ZTVN10__cxxabiv117__class_type_infoE = external global i8*
+// CHECK: @_ZTS1A = constant [3 x i8] c"1A\00", align 1
+// CHECK: @_ZTI1A = constant { i8*, i8* } { i8* getelementptr inbounds (i8, i8* bitcast (i8** @_ZTVN10__cxxabiv117__class_type_infoE to i8*), i32 8), i8* getelementptr inbounds ([3 x i8], [3 x i8]* @_ZTS1A, i32 0, i32 0) }, align 8
+// CHECK: @_ZTVN10__cxxabiv120__si_class_type_infoE = external global i8*
+// CHECK: @_ZTS1B = constant [3 x i8] c"1B\00", align 1
+// CHECK: @_ZTI1B = constant { i8*, i8*, i8* } { i8* getelementptr inbounds (i8, i8* bitcast (i8** @_ZTVN10__cxxabiv120__si_class_type_infoE to i8*), i32 8), i8* getelementptr inbounds ([3 x i8], [3 x i8]* @_ZTS1B, i32 0, i32 0), i8* bitcast ({ i8*, i8* }* @_ZTI1A to i8*) }, align 8
+// CHECK: @_ZTI1A.rtti_proxy = hidden unnamed_addr constant { i8*, i8* }* @_ZTI1A, comdat
+// CHECK: @_ZTI1B.rtti_proxy = hidden unnamed_addr constant { i8*, i8*, i8* }* @_ZTI1B, comdat
+
+// CHECK:      define dereferenceable(16) %"class.std::type_info"* @_Z11getTypeInfov() local_unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret %"class.std::type_info"* bitcast ({ i8*, i8* }* @_ZTI1A to %"class.std::type_info"*)
+// CHECK-NEXT: }
+
+// CHECK:      define i8* @_Z7getNamev() local_unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   ret i8* getelementptr inbounds ([3 x i8], [3 x i8]* @_ZTS1A, i64 0, i64 0)
+// CHECK-NEXT: }
+
+// CHECK:      define i1 @_Z5equalP1A(%class.A* readonly %a) local_unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   [[isnull:%[0-9]+]] = icmp eq %class.A* %a, null
+// CHECK-NEXT:   br i1 [[isnull]], label %[[bad_typeid:[a-z0-9._]+]], label %[[end:[a-z0-9.+]+]]
+// CHECK:      [[bad_typeid]]:
+// CHECK-NEXT:   tail call void @__cxa_bad_typeid()
+// CHECK-NEXT:   unreachable
+// CHECK:      [[end]]:
+// CHECK-NEXT:   [[type_info_ptr3:%[0-9]+]] = bitcast %class.A* %a to i8**
+// CHECK-NEXT:   [[vtable:%[a-z0-9]+]] = load i8*, i8** [[type_info_ptr3]]
+// CHECK-NEXT:   [[type_info_ptr:%[0-9]+]] = tail call i8* @llvm.load.relative.i32(i8* [[vtable]], i32 -4)
+// CHECK-NEXT:   [[type_info_ptr2:%[0-9]+]] = bitcast i8* [[type_info_ptr]] to %"class.std::type_info"**
+// CHECK-NEXT:   [[type_info_ptr:%[0-9]+]] = load %"class.std::type_info"*, %"class.std::type_info"** [[type_info_ptr2]], align 8
+// CHECK-NEXT:   [[name_ptr:%[a-z0-9._]+]] = getelementptr inbounds %"class.std::type_info", %"class.std::type_info"* [[type_info_ptr]], i64 0, i32 1
+// CHECK-NEXT:   [[name:%[0-9]+]] = load i8*, i8** [[name_ptr]], align 8
+// CHECK-NEXT:   [[eq:%[a-z0-9.]+]] = icmp eq i8* [[name]], getelementptr inbounds ([3 x i8], [3 x i8]* @_ZTS1B, i64 0, i64 0)
+// CHECK-NEXT:   ret i1 [[eq]]
+// CHECK-NEXT: }
+
+#include "../typeinfo"
+
+class A {
+public:
+  virtual void foo();
+};
+
+class B : public A {
+public:
+  void foo() override;
+};
+
+void A::foo() {}
+void B::foo() {}
+
+const auto &getTypeInfo() {
+  return typeid(A);
+}
+
+const char *getName() {
+  return typeid(A).name();
+}
+
+bool equal(A *a) {
+  return typeid(B) == typeid(*a);
+}

clang/test/CodeGenCXX/Fuchsia/vbase-offset.cpp

+// Check that the pointer adjustment from the virtual base offset is loaded as a
+// 32-bit int.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK-LABEL: @_ZTv0_n12_N7Derived1fEi(
+// CHECK-NEXT:  entry:
+// CHECK:        [[this:%.+]] = bitcast %class.Derived* %this1 to i8*
+// CHECK-NEXT:   [[this2:%.+]] = bitcast i8* [[this]] to i8**
+// CHECK-NEXT:   [[vtable:%.+]] = load i8*, i8** [[this2]], align 8
+// CHECK-NEXT:   [[vbase_offset_ptr:%.+]] = getelementptr inbounds i8, i8* [[vtable]], i64 -12
+// CHECK-NEXT:   [[vbase_offset_ptr2:%.+]] = bitcast i8* [[vbase_offset_ptr]] to i32*
+// CHECK-NEXT:   [[vbase_offset:%.+]] = load i32, i32* [[vbase_offset_ptr2]], align 4
+// CHECK-NEXT:   [[adj_this:%.+]] = getelementptr inbounds i8, i8* [[this]], i32 [[vbase_offset]]
+// CHECK-NEXT:   [[adj_this2:%.+]] = bitcast i8* [[adj_this]] to %class.Derived*
+// CHECK:        [[call:%.+]] = tail call i32 @_ZN7Derived1fEi(%class.Derived* [[adj_this2]], i32 {{.*}})
+// CHECK:        ret i32 [[call]]
+
+class Base {
+public:
+  virtual int f(int x);
+
+private:
+  long x;
+};
+
+class Derived : public virtual Base {
+public:
+  virtual int f(int x);
+
+private:
+  long y;
+};
+
+int Base::f(int x) { return x + 1; }
+int Derived::f(int x) { return x + 2; }

clang/test/CodeGenCXX/Fuchsia/virtual-function-call.cpp

+// Check that we call llvm.load.relative() on a vtable function call.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -O3 -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK:      define void @_Z5A_fooP1A(%class.A* %a) local_unnamed_addr
+// CHECK-NEXT: entry:
+// CHECK-NEXT:   [[this:%[0-9]+]] = bitcast %class.A* %a to i8**
+// CHECK-NEXT:   %vtable1 = load i8*, i8** [[this]]
+// CHECK-NEXT:   [[func_ptr:%[0-9]+]] = tail call i8* @llvm.load.relative.i32(i8* %vtable1, i32 0)
+// CHECK-NEXT:   [[func:%[0-9]+]] = bitcast i8* [[func_ptr]] to void (%class.A*)*
+// CHECK-NEXT:   tail call void [[func]](%class.A* %a)
+// CHECK-NEXT:   ret void
+// CHECK-NEXT: }
+
+class A {
+public:
+  virtual void foo();
+};
+
+void A_foo(A *a) {
+  a->foo();
+}

clang/test/CodeGenCXX/Fuchsia/vtable-hidden-when-in-comdat.cpp

+// Check that a vtable is made hidden instead of private if the original vtable
+// is not dso_local. The vtable will need to be hidden and not private so it can
+// be used as acomdat key signature.
+
+// RUN: %clang_cc1 %s -triple=aarch64-unknown-fuchsia -S -o - -emit-llvm -fexperimental-relative-c++-abi-vtables | FileCheck %s
+
+// CHECK: @_ZTV1B.local = linkonce_odr hidden unnamed_addr constant
+// CHECK: @_ZTV1B = linkonce_odr unnamed_addr alias { [3 x i32] }, { [3 x i32] }* @_ZTV1B.local
+
+// The VTable will be in a comdat here since it has no key function.
+class B {
+ public:
+  inline virtual void func() {}
+};
+
+// This is here just to manifest the vtable for B.
+void func() {
+  B b;
+}

clang/test/CodeGenCXX/constructor-destructor-return-this.cpp

 //RUN: %clang_cc1 %s -emit-llvm -o - -triple=i686-unknown-linux | FileCheck --check-prefix=CHECKGEN %s
 //RUN: %clang_cc1 %s -emit-llvm -o - -triple=thumbv7-apple-ios6.0 -target-abi apcs-gnu | FileCheck --check-prefix=CHECKARM %s
 //RUN: %clang_cc1 %s -emit-llvm -o - -triple=thumbv7-apple-ios5.0 -target-abi apcs-gnu | FileCheck --check-prefix=CHECKIOS5 %s
 //RUN: %clang_cc1 %s -emit-llvm -o - -triple=wasm32-unknown-unknown \
 //RUN:   | FileCheck --check-prefix=CHECKARM %s
 //RUN: %clang_cc1 %s -emit-llvm -o - -triple=x86_64-unknown-fuchsia | FileCheck --check-prefix=CHECKFUCHSIA %s
 //RUN: %clang_cc1 %s -emit-llvm -o - -triple=aarch64-unknown-fuchsia | FileCheck --check-prefix=CHECKFUCHSIA %s
+//RUN: %clang_cc1 %s -emit-llvm -o - -triple=x86_64-unknown-fuchsia -fexperimental-relative-c++-abi-vtables | FileCheck --check-prefix=CHECKFUCHSIA-RELATIVE-ABI %s
+//RUN: %clang_cc1 %s -emit-llvm -o - -triple=aarch64-unknown-fuchsia -fexperimental-relative-c++-abi-vtables | FileCheck --check-prefix=CHECKFUCHSIA-RELATIVE-ABI %s
 //RUN: %clang_cc1 %s -emit-llvm -o - -triple=i386-pc-win32 -fno-rtti | FileCheck --check-prefix=CHECKMS %s
 // FIXME: these tests crash on the bots when run with -triple=x86_64-pc-win32
 
 // Make sure we attach the 'returned' attribute to the 'this' parameter of
 // constructors and destructors which return this (and only these cases)
 
 class A {
 public:
void test_destructor() {
   E* e2 = gete();
   e2->~E();
 }
 
 // CHECKARM-LABEL,CHECKFUCHSIA-LABEL: define void @_Z15test_destructorv()
 
 // Verify that virtual calls to destructors are not marked with a 'returned'
 // this parameter at the call site...
-// CHECKARM,CHECKFUCHSIA: [[VFN:%.*]] = getelementptr inbounds %class.E* (%class.E*)*, %class.E* (%class.E*)**
-// CHECKARM,CHECKFUCHSIA: [[THUNK:%.*]] = load %class.E* (%class.E*)*, %class.E* (%class.E*)** [[VFN]]
+// CHECKARM: [[VFN:%.*]] = getelementptr inbounds %class.E* (%class.E*)*, %class.E* (%class.E*)**
+// CHECKARM: [[THUNK:%.*]] = load %class.E* (%class.E*)*, %class.E* (%class.E*)** [[VFN]]
+// CHECKFUCHSIA: [[VFN:%.*]] = getelementptr inbounds %class.E* (%class.E*)*, %class.E* (%class.E*)**
+// CHECKFUCHSIA: [[THUNK:%.*]] = load %class.E* (%class.E*)*, %class.E* (%class.E*)** [[VFN]]
+// CHECKFUCHSIA-RELATIVE-ABI: [[VTABLE:%.+]] = bitcast %class.E* (%class.E*)**
+// CHECKFUCHSIA-RELATIVE-ABI: [[THUNK_PTR:%.+]] = call i8* @llvm.load.relative.i32(i8* [[VTABLE]], i32 0)
+// CHECKFUCHSIA-RELATIVE-ABI: [[THUNK:%.+]] = bitcast i8* [[THUNK_PTR]] to %class.E* (%class.E*)*
 // CHECKARM,CHECKFUCHSIA: call %class.E* [[THUNK]](%class.E* %
 
 // ...but static calls create declarations with 'returned' this
 // CHECKARM,CHECKFUCHSIA: {{%.*}} = call %class.E* @_ZN1ED1Ev(%class.E* %
 
 // CHECKARM,CHECKFUCHSIA: declare %class.E* @_ZN1ED1Ev(%class.E* returned)

libcxxabi/CMakeLists.txt

 endif()
 
 set(LIBCXXABI_LIBCXX_PATH "${LIBCXXABI_LIBCXX_PATH}" CACHE PATH
     "Specify path to libc++ source." FORCE)
 
 option(LIBCXXABI_HERMETIC_STATIC_LIBRARY
   "Do not export any symbols from the static library." OFF)
 
+option(LIBCXXABI_RELATIVE_VTABLES_ABI_ENABLED
+  "Use different configurations of libcxxabi functions supported for the \
+  Relative VTables ABI, which replaces vtable function pointers with offsets \
+  between those virtual functions and the vtable itself." OFF)
+
 #===============================================================================
 # Configure System
 #===============================================================================
 
 # Add path for custom modules
 set(CMAKE_MODULE_PATH
   "${CMAKE_CURRENT_SOURCE_DIR}/cmake"
   "${CMAKE_CURRENT_SOURCE_DIR}/cmake/Modules"
if (LIBCXXABI_HAS_PTHREAD_LIB)
     add_definitions(-D_LIBCXXABI_LINK_PTHREAD_LIB)
   endif()
 endif()
 
 string(REPLACE ";" " " LIBCXXABI_CXX_FLAGS "${LIBCXXABI_CXX_FLAGS}")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${LIBCXXABI_CXX_FLAGS}")
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${LIBCXXABI_C_FLAGS}")
 
+if (LIBCXXABI_RELATIVE_VTABLES_ABI_ENABLED)
+  add_definitions(-D_LIBCXXABI_RELATIVE_VTABLES_ABI_ENABLED)
+endif()
+
 #===============================================================================
 # Setup Source Code
 #===============================================================================
 
 set(LIBCXXABI_LIBUNWIND_INCLUDES "${LIBCXXABI_LIBUNWIND_INCLUDES}" CACHE PATH
     "Specify path to libunwind includes." FORCE)
 set(LIBCXXABI_LIBUNWIND_PATH "${LIBCXXABI_LIBUNWIND_PATH}" CACHE PATH
     "Specify path to libunwind source." FORCE)

libcxxabi/src/private_typeinfo.cpp

 
 extern "C" _LIBCXXABI_FUNC_VIS void *
 __dynamic_cast(const void *static_ptr, const __class_type_info *static_type,
                const __class_type_info *dst_type,
                std::ptrdiff_t src2dst_offset) {
     // Possible future optimization:  Take advantage of src2dst_offset
 
     // Get (dynamic_ptr, dynamic_type) from static_ptr
-    void **vtable = *static_cast<void ** const *>(static_ptr);
+#ifdef _LIBCXXABI_RELATIVE_VTABLES_ABI_ENABLED
+    // The vtable address will point to the first virtual function, which is 8
+    // bytes after the start of the vtable (4 for the offset from top + 4 for the typeinfo component).
+    uint32_t* vtable = *reinterpret_cast<uint32_t* const*>(static_ptr);
+    uint32_t offset_to_derived = vtable[-2];
+
+    const void* dynamic_ptr = static_cast<const char*>(static_ptr) + offset_to_derived;
+
+    // The typeinfo component is now a relative offset to a proxy.
+    int32_t offset_to_ti_proxy = *reinterpret_cast<int32_t *>(vtable - 4);
+    uint8_t *ptr_to_ti_proxy = vtable + offset_to_ti_proxy;
+    const __class_type_info* dynamic_type = *(reinterpret_cast<__class_type_info **>(ptr_to_ti_proxy));
+#else
+    void** vtable = *static_cast<void** const*>(static_ptr);
     ptrdiff_t offset_to_derived = reinterpret_cast<ptrdiff_t>(vtable[-2]);
-    const void* dynamic_ptr = static_cast<const char*>(static_ptr) + offset_to_derived;
-    const __class_type_info* dynamic_type = static_cast<const __class_type_info*>(vtable[-1]);
+    const void* dynamic_ptr =
+        static_cast<const char*>(static_ptr) + offset_to_derived;
+    const __class_type_info* dynamic_type =
+        static_cast<const __class_type_info*>(vtable[-1]);
+#endif
 
     // Initialize answer to nullptr.  This will be changed from the search
     //    results if a non-null answer is found.  Regardless, this is what will
     //    be returned.
     const void* dst_ptr = 0;
     // Initialize info struct for this search.
     __dynamic_cast_info info = {dst_type, static_ptr, static_type, src2dst_offset, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,};
 

llvm/lib/IR/Constants.cpp

if (CE->getOpcode() == Instruction::Sub) {
         // function.  This is a common idiom when creating a table for the
         // indirect goto extension, so we handle it efficiently here.
         if (isa<BlockAddress>(LHSOp0) && isa<BlockAddress>(RHSOp0) &&
             cast<BlockAddress>(LHSOp0)->getFunction() ==
                 cast<BlockAddress>(RHSOp0)->getFunction())
           return false;
 
         // Relative pointers do not need to be dynamically relocated.
-        if (auto *LHSGV = dyn_cast<GlobalValue>(LHSOp0->stripPointerCasts()))
-          if (auto *RHSGV = dyn_cast<GlobalValue>(RHSOp0->stripPointerCasts()))
+        if (auto *LHSGV =
+                dyn_cast<GlobalValue>(LHSOp0->stripInBoundsConstantOffsets()))
+          if (auto *RHSGV =
+                  dyn_cast<GlobalValue>(RHSOp0->stripInBoundsConstantOffsets()))
             if (LHSGV->isDSOLocal() && RHSGV->isDSOLocal())
               return false;
       }
     }
   }
 
   bool Result = false;
   for (unsigned i = 0, e = getNumOperands(); i != e; ++i)

llvm/test/CodeGen/X86/relptr-rodata.ll

 ; CHECK: relro1:
 ; CHECK: .long default-relro1
 @relro1 = hidden constant i32 trunc (i64 sub (i64 ptrtoint (i8* @default to i64), i64 ptrtoint (i32* @relro1 to i64)) to i32)
 
 ; CHECK: .section .data.rel.ro.relro2
 ; CHECK: relro2:
 ; CHECK: .long hidden-relro2
 @relro2 = constant i32 trunc (i64 sub (i64 ptrtoint (i8* @hidden to i64), i64 ptrtoint (i32* @relro2 to i64)) to i32)
+
+; CHECK:      .section .rodata._ZTV4Test
+; CHECK-NEXT: .globl _ZTV4Test
+; CHECK:      _ZTV4Test
+; CHECK:      .long 0
+; CHECK:      .long (_ZTI4Test.rtti_proxy-_ZTV4Test)-4
+; CHECK:      .long (_ZN4TestD1Ev.stub@PLT-_ZTV4Test)-8
+
+@_ZTI4Test = external global i8
+@_ZTI4Test.rtti_proxy = hidden unnamed_addr constant i8* @_ZTI4Test
+
+declare void @_ZN4TestD2Ev()
+define hidden void @_ZN4TestD1Ev.stub() unnamed_addr {
+entry:
+  tail call void @_ZN4TestD2Ev()
+  ret void
+}
+
+@_ZTV4Test = dso_local unnamed_addr constant { [3 x i32] } {
+  [3 x i32] [
+    i32 0,
+    i32 trunc (i64 sub (i64 ptrtoint (i8** @_ZTI4Test.rtti_proxy to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV4Test, i32 0, i32 0, i32 1) to i64)) to i32),
+    i32 trunc (i64 sub (i64 ptrtoint (void ()* @_ZN4TestD1Ev.stub to i64), i64 ptrtoint (i32* getelementptr inbounds ({ [3 x i32] }, { [3 x i32] }* @_ZTV4Test, i32 0, i32 0, i32 2) to i64)) to i32)
+  ] }, align 4