[RuntimeDyld] Added support for relocation of indirect functions

[MoritzS]

Summary

In ELF, symbols of type STT_GNU_IFUNC need to be resolved by calling the
function at the symbol's address. This is implemented by adding special
stubs for all symbols of that type.

[lhames]

This is really cool.

Does the Indirect flag really need to go onto JITSymbolFlags? My mental model of ifuncs (very limited -- I haven't used them myself) is "self-updating PLT" -- its ifunc-ness is opaque to callers right?

[MoritzS]

In D105465#3653761, @lhames wrote:

Does the Indirect flag really need to go onto JITSymbolFlags? My mental model of ifuncs (very limited -- I haven't used them myself) is "self-updating PLT" -- its ifunc-ness is opaque to callers right?

Yes, you are right, ifuncs are completely opaque to callers, so we probably shouldn't add that flag. I will change that.

[MoritzS]

Removed Indirect from JITSymbolFlags

[MoritzS]

Fixed typos in comments

[MoritzS]

Rebased onto main

[MoritzS]

Rebased again because of failing tests on main

[MoritzS]

@lhames do you have any other comments? Can I merge this?

[MoritzS]

Rebased onto main, let tests run again

[MoritzS]

Fixed a failing test: In ELFObjectFile I added a case to set the symbol type to
ST_Function also for ifunc symbols. As far as I know, ifunc symbols can only be
functions, but even objdump from binutils doesn't print those symbols with the
function type. I removed that case again, ifunc symbols now have type ST_Other.

[MoritzS]

Fixed an unchecked Expected<T> which lead to several test failures. All tests should run now, hopefully.

[MoritzS]

Use llvm::cantFail instead of just returning

[MoritzS]

The only failing "test" is clang-format. Should I reformat the complete files? Because my additions are formatted correctly, but the rest of the file is not.

[lhames]

LGTM, and sorry for the very delayed review!

What are your blockers to moving over to ORC/JITLink these days? This is excellent work -- if there's something that I can do to unblock you so that you can work on the new JIT linker I'd be keen to know. ;)

[lhames]

The only failing "test" is clang-format. Should I reformat the complete files? Because my additions are formatted correctly, but the rest of the file is not.

Only your additions should be reformatted. Reformatting the rest of the files will obscure the commit history. It's better to let them be updated to the correct format over time as patches come in. (Though my hope is that we can kill RuntimeDyld in the not too distant future and just move everyone over to JITLink)

llvm/include/llvm/Object/ELFObjectFile.h

ELFObjectFile<ELFT>::getSymbolType(DataRefImpl Symb) const {
   switch ((*SymOrErr)->getType()) {
   case ELF::STT_NOTYPE:
     return SymbolRef::ST_Unknown;
   case ELF::STT_SECTION:
     return SymbolRef::ST_Debug;
   case ELF::STT_FILE:
     return SymbolRef::ST_File;
   case ELF::STT_FUNC:
+  case ELF::STT_GNU_IFUNC:
     return SymbolRef::ST_Function;
   case ELF::STT_OBJECT:
   case ELF::STT_COMMON:
     return SymbolRef::ST_Data;
   case ELF::STT_TLS:
   default:
     return SymbolRef::ST_Other;
   }
if (ESym->st_shndx == ELF::SHN_UNDEF)
     Result |= SymbolRef::SF_Undefined;
 
   if (ESym->getType() == ELF::STT_COMMON || ESym->st_shndx == ELF::SHN_COMMON)
     Result |= SymbolRef::SF_Common;
 
   if (isExportedToOtherDSO(ESym))
     Result |= SymbolRef::SF_Exported;
 
+  if (ESym->getType() == ELF::STT_GNU_IFUNC)
+    Result |= SymbolRef::SF_Indirect;
+
   if (ESym->getVisibility() == ELF::STV_HIDDEN)
     Result |= SymbolRef::SF_Hidden;
 
   return Result;
 }
 
 template <class ELFT>
 Expected<section_iterator>

llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyld.cpp

if (*FlagsOrErr & SymbolRef::SF_Absolute &&
         return AddrOrErr.takeError();
 
       unsigned SectionID = AbsoluteSymbolSection;
 
       LLVM_DEBUG(dbgs() << "\tType: " << SymType << " (absolute) Name: " << Name
                         << " SID: " << SectionID
                         << " Offset: " << format("%p", (uintptr_t)Addr)
                         << " flags: " << *FlagsOrErr << "\n");
-      if (!Name.empty()) // Skip absolute symbol relocations.
-        GlobalSymbolTable[Name] =
-            SymbolTableEntry(SectionID, Addr, *JITSymFlags);
+      // Skip absolute symbol relocations.
+      if (!Name.empty()) {
+        auto Result = GlobalSymbolTable.insert_or_assign(
+            Name, SymbolTableEntry(SectionID, Addr, *JITSymFlags));
+        processNewSymbol(*I, Result.first->getValue());
+      }
     } else if (SymType == object::SymbolRef::ST_Function ||
                SymType == object::SymbolRef::ST_Data ||
                SymType == object::SymbolRef::ST_Unknown ||
                SymType == object::SymbolRef::ST_Other) {
 
       section_iterator SI = Obj.section_end();
       if (auto SIOrErr = I->getSection())
         SI = *SIOrErr;
if (*FlagsOrErr & SymbolRef::SF_Absolute &&
         SectionID = *SectionIDOrErr;
       else
         return SectionIDOrErr.takeError();
 
       LLVM_DEBUG(dbgs() << "\tType: " << SymType << " Name: " << Name
                         << " SID: " << SectionID
                         << " Offset: " << format("%p", (uintptr_t)SectOffset)
                         << " flags: " << *FlagsOrErr << "\n");
-      if (!Name.empty()) // Skip absolute symbol relocations
-        GlobalSymbolTable[Name] =
-            SymbolTableEntry(SectionID, SectOffset, *JITSymFlags);
+      // Skip absolute symbol relocations.
+      if (!Name.empty()) {
+        auto Result = GlobalSymbolTable.insert_or_assign(
+            Name, SymbolTableEntry(SectionID, SectOffset, *JITSymFlags));
+        processNewSymbol(*I, Result.first->getValue());
+      }
     }
   }
 
   // Allocate common symbols
   if (auto Err = emitCommonSymbols(Obj, CommonSymbolsToAllocate, CommonSize,
                                    CommonAlign))
     return std::move(Err);
 
if (*FlagsOrErr & SymbolRef::SF_Common) {
       CommonSize = alignTo(CommonSize, Align) + Size;
     }
   }
   if (CommonSize != 0) {
     RWSectionSizes.push_back(CommonSize);
     RWDataAlign = std::max(RWDataAlign, CommonAlign);
   }
 
+  if (!CodeSectionSizes.empty()) {
+    // Add 64 bytes for a potential IFunc resolver stub
+    CodeSectionSizes.push_back(64);
+  }
+
   // Compute the required allocation space for each different type of sections
   // (code, read-only data, read-write data) assuming that all sections are
   // allocated with the max alignment. Note that we cannot compute with the
   // individual alignments of the sections, because then the required size
   // depends on the order, in which the sections are allocated.
   CodeSize = computeAllocationSizeForSections(CodeSectionSizes, CodeAlign);
   RODataSize = computeAllocationSizeForSections(ROSectionSizes, RODataAlign);
   RWDataSize = computeAllocationSizeForSections(RWSectionSizes, RWDataAlign);

llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.h

private:
   // When a module is loaded we save the SectionID of the EH frame section
   // in a table until we receive a request to register all unregistered
   // EH frame sections with the memory manager.
   SmallVector<SID, 2> UnregisteredEHFrameSections;
 
   // Map between GOT relocation value and corresponding GOT offset
   std::map<RelocationValueRef, uint64_t> GOTOffsetMap;
 
+  /// The ID of the current IFunc stub section
+  unsigned IFuncStubSectionID = 0;
+  /// The current offset into the IFunc stub section
+  uint64_t IFuncStubOffset = 0;
+
+  /// A IFunc stub and its original symbol
+  struct IFuncStub {
+    /// The offset of this stub in the IFunc stub section
+    uint64_t StubOffset;
+    /// The symbol table entry of the original symbol
+    SymbolTableEntry OriginalSymbol;
+  };
+
+  /// The IFunc stubs
+  SmallVector<IFuncStub, 2> IFuncStubs;
+
+  /// Create the code for the IFunc resolver at the given address. This code
+  /// works together with the stubs created in createIFuncStub() to call the
+  /// resolver function and then jump to the real function address.
+  /// It must not be larger than 64B.
+  void createIFuncResolver(uint8_t *Addr) const;
+  /// Create the code for an IFunc stub for the IFunc that is defined in
+  /// section IFuncSectionID at offset IFuncOffset. The IFunc resolver created
+  /// by createIFuncResolver() is defined in the section IFuncStubSectionID at
+  /// offset IFuncResolverOffset. The code should be written into the section
+  /// with the id IFuncStubSectionID at the offset IFuncStubOffset.
+  void createIFuncStub(unsigned IFuncStubSectionID,
+                       uint64_t IFuncResolverOffset, uint64_t IFuncStubOffset,
+                       unsigned IFuncSectionID, uint64_t IFuncOffset);
+  /// Return the maximum size of a stub created by createIFuncStub()
+  unsigned getMaxIFuncStubSize() const;
+
+  void processNewSymbol(const SymbolRef &ObjSymbol,
+                        SymbolTableEntry &Entry) override;
   bool relocationNeedsGot(const RelocationRef &R) const override;
   bool relocationNeedsStub(const RelocationRef &R) const override;
 
   // Process a GOTTPOFF TLS relocation for x86-64
   // NOLINTNEXTLINE(readability-identifier-naming)
   void processX86_64GOTTPOFFRelocation(unsigned SectionID, uint64_t Offset,
                                        RelocationValueRef Value,
                                        int64_t Addend);

llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp

 }
 
 RelocationEntry RuntimeDyldELF::computeGOTOffsetRE(uint64_t GOTOffset,
                                                    uint64_t SymbolOffset,
                                                    uint32_t Type) {
   return RelocationEntry(GOTSectionID, GOTOffset, Type, SymbolOffset);
 }
 
+void RuntimeDyldELF::processNewSymbol(const SymbolRef &ObjSymbol, SymbolTableEntry& Symbol) {
+  // This should never return an error as `processNewSymbol` wouldn't have been
+  // called if getFlags() returned an error before.
+  auto ObjSymbolFlags = ObjSymbol.getFlags().get();
+  if (ObjSymbolFlags & SymbolRef::SF_Indirect) {
+    if (IFuncStubSectionID == 0) {
+      // Create a dummy section for the ifunc stubs. It will be actually
+      // allocated in finalizeLoad() below.
+      IFuncStubSectionID = Sections.size();
+      Sections.push_back(
+          SectionEntry(".text.__llvm_IFuncStubs", nullptr, 0, 0, 0));
+      // First 64B are reserverd for the IFunc resolver
+      IFuncStubOffset = 64;
+    }
+
+    IFuncStubs.push_back(IFuncStub{IFuncStubOffset, Symbol});
+    // Modify the symbol so that it points to the ifunc stub instead of to the
+    // resolver function.
+    Symbol = SymbolTableEntry(IFuncStubSectionID, IFuncStubOffset,
+                              Symbol.getFlags());
+    IFuncStubOffset += getMaxIFuncStubSize();
+  }
+}
+
 Error RuntimeDyldELF::finalizeLoad(const ObjectFile &Obj,
                                   ObjSectionToIDMap &SectionMap) {
   if (IsMipsO32ABI)
     if (!PendingRelocs.empty())
       return make_error<RuntimeDyldError>("Can't find matching LO16 reloc");
 
+  // Create the IFunc stubs if necessary. This must be done before processing
+  // the GOT entries, as the IFunc stubs may create some.
+  if (IFuncStubSectionID != 0) {
+    uint8_t *IFuncStubsAddr = MemMgr.allocateCodeSection(
+        IFuncStubOffset, 1, IFuncStubSectionID, ".text.__llvm_IFuncStubs");
+    if (!IFuncStubsAddr)
+      return make_error<RuntimeDyldError>(
+          "Unable to allocate memory for IFunc stubs!");
+    Sections[IFuncStubSectionID] =
+        SectionEntry(".text.__llvm_IFuncStubs", IFuncStubsAddr, IFuncStubOffset,
+                     IFuncStubOffset, 0);
+
+    createIFuncResolver(IFuncStubsAddr);
+
+    LLVM_DEBUG(dbgs() << "Creating IFunc stubs SectionID: "
+                      << IFuncStubSectionID << " Addr: "
+                      << Sections[IFuncStubSectionID].getAddress() << '\n');
+    for (auto &IFuncStub : IFuncStubs) {
+      auto &Symbol = IFuncStub.OriginalSymbol;
+      LLVM_DEBUG(dbgs() << "\tSectionID: " << Symbol.getSectionID()
+                        << " Offset: " << format("%p", Symbol.getOffset())
+                        << " IFuncStubOffset: "
+                        << format("%p\n", IFuncStub.StubOffset));
+      createIFuncStub(IFuncStubSectionID, 0, IFuncStub.StubOffset,
+                      Symbol.getSectionID(), Symbol.getOffset());
+    }
+
+    IFuncStubSectionID = 0;
+    IFuncStubOffset = 0;
+    IFuncStubs.clear();
+  }
+
   // If necessary, allocate the global offset table
   if (GOTSectionID != 0) {
     // Allocate memory for the section
     size_t TotalSize = CurrentGOTIndex * getGOTEntrySize();
     uint8_t *Addr = MemMgr.allocateDataSection(TotalSize, getGOTEntrySize(),
-                                                GOTSectionID, ".got", false);
     if (!Addr)
       return make_error<RuntimeDyldError>("Unable to allocate memory for GOT!");
 
     Sections[GOTSectionID] =
         SectionEntry(".got", Addr, TotalSize, TotalSize, 0);
 
     // For now, initialize all GOT entries to zero.  We'll fill them in as
     // needed when GOT-based relocations are applied.
     memset(Addr, 0, TotalSize);
     if (IsMipsN32ABI || IsMipsN64ABI) {
       // To correctly resolve Mips GOT relocations, we need a mapping from
       // object's sections to GOTs.
       for (section_iterator SI = Obj.section_begin(), SE = Obj.section_end();
            SI != SE; ++SI) {
         if (SI->relocation_begin() != SI->relocation_end()) {
           Expected<section_iterator> RelSecOrErr = SI->getRelocatedSection();
           if (!RelSecOrErr)
             return make_error<RuntimeDyldError>(
                 toString(RelSecOrErr.takeError()));
 
           section_iterator RelocatedSection = *RelSecOrErr;
           ObjSectionToIDMap::iterator i = SectionMap.find(*RelocatedSection);
-          assert (i != SectionMap.end());
+          assert(i != SectionMap.end());
           SectionToGOTMap[i->second] = GOTSectionID;
         }
       }
       GOTSymbolOffsets.clear();
     }
   }
 
   // Look for and record the EH frame section.
Error RuntimeDyldELF::finalizeLoad(const ObjectFile &Obj,
 
   return Error::success();
 }
 
 bool RuntimeDyldELF::isCompatibleFile(const object::ObjectFile &Obj) const {
   return Obj.isELF();
 }
 
+void RuntimeDyldELF::createIFuncResolver(uint8_t *Addr) const {
+  if (Arch == Triple::x86_64) {
+    // The adddres of the GOT1 entry is in %r11, the GOT2 entry is in %r11+8
+    // (see createIFuncStub() for details)
+    // The following code first saves all registers that contain the original
+    // function arguments as those registers are not saved by the resolver
+    // function. %r11 is saved as well so that the GOT2 entry can be updated
+    // afterwards. Then it calls the actual IFunc resolver function whose
+    // address is stored in GOT2. After the resolver function returns, all
+    // saved registers are restored and the return value is written to GOT1.
+    // Finally, jump to the now resolved function.
+    // clang-format off
+    const uint8_t StubCode[] = {
+        0x57,                   // push %rdi
+        0x56,                   // push %rsi
+        0x52,                   // push %rdx
+        0x51,                   // push %rcx
+        0x41, 0x50,             // push %r8
+        0x41, 0x51,             // push %r9
+        0x41, 0x53,             // push %r11
+        0x41, 0xff, 0x53, 0x08, // call *0x8(%r11)
+        0x41, 0x5b,             // pop %r11
+        0x41, 0x59,             // pop %r9
+        0x41, 0x58,             // pop %r8
+        0x59,                   // pop %rcx
+        0x5a,                   // pop %rdx
+        0x5e,                   // pop %rsi
+        0x5f,                   // pop %rdi
+        0x49, 0x89, 0x03,       // mov %rax,(%r11)
+        0xff, 0xe0              // jmp *%rax
+    };
+    // clang-format on
+    static_assert(sizeof(StubCode) <= 64,
+                  "maximum size of the IFunc resolver is 64B");
+    memcpy(Addr, StubCode, sizeof(StubCode));
+  } else {
+    report_fatal_error(
+        "IFunc resolver is not supported for target architecture");
+  }
+}
+
+void RuntimeDyldELF::createIFuncStub(unsigned IFuncStubSectionID,
+                                     uint64_t IFuncResolverOffset,
+                                     uint64_t IFuncStubOffset,
+                                     unsigned IFuncSectionID,
+                                     uint64_t IFuncOffset) {
+  auto &IFuncStubSection = Sections[IFuncStubSectionID];
+  auto *Addr = IFuncStubSection.getAddressWithOffset(IFuncStubOffset);
+
+  if (Arch == Triple::x86_64) {
+    // The first instruction loads a PC-relative address into %r11 which is a
+    // GOT entry for this stub. This initially contains the address to the
+    // IFunc resolver. We can use %r11 here as it's caller saved but not used
+    // to pass any arguments. In fact, x86_64 ABI even suggests using %r11 for
+    // code in the PLT. The IFunc resolver will use %r11 to update the GOT
+    // entry.
+    //
+    // The next instruction just jumps to the address contained in the GOT
+    // entry. As mentioned above, we do this two-step jump by first setting
+    // %r11 so that the IFunc resolver has access to it.
+    //
+    // The IFunc resolver of course also needs to know the actual address of
+    // the actual IFunc resolver function. This will be stored in a GOT entry
+    // right next to the first one for this stub. So, the IFunc resolver will
+    // be able to call it with %r11+8.
+    //
+    // In total, two adjacent GOT entries (+relocation) and one additional
+    // relocation are required:
+    // GOT1: Address of the IFunc resolver.
+    // GOT2: Address of the IFunc resolver function.
+    // IFuncStubOffset+3: 32-bit PC-relative address of GOT1.
+    uint64_t GOT1 = allocateGOTEntries(2);
+    uint64_t GOT2 = GOT1 + getGOTEntrySize();
+
+    RelocationEntry RE1(GOTSectionID, GOT1, ELF::R_X86_64_64,
+                        IFuncResolverOffset, {});
+    addRelocationForSection(RE1, IFuncStubSectionID);
+    RelocationEntry RE2(GOTSectionID, GOT2, ELF::R_X86_64_64, IFuncOffset, {});
+    addRelocationForSection(RE2, IFuncSectionID);
+
+    const uint8_t StubCode[] = {
+        0x4c, 0x8d, 0x1d, 0x00, 0x00, 0x00, 0x00, // leaq 0x0(%rip),%r11
+        0x41, 0xff, 0x23                          // jmpq *(%r11)
+    };
+    assert(sizeof(StubCode) <= getMaxIFuncStubSize() &&
+           "IFunc stub size must not exceed getMaxIFuncStubSize()");
+    memcpy(Addr, StubCode, sizeof(StubCode));
+
+    // The PC-relative value starts 4 bytes from the end of the leaq
+    // instruction, so the addend is -4.
+    resolveGOTOffsetRelocation(IFuncStubSectionID, IFuncStubOffset + 3,
+                               GOT1 - 4, ELF::R_X86_64_PC32);
+  } else {
+    report_fatal_error("IFunc stub is not supported for target architecture");
+  }
+}
+
+unsigned RuntimeDyldELF::getMaxIFuncStubSize() const {
+  if (Arch == Triple::x86_64) {
+    return 10;
+  }
+  return 0;
+}
+
 bool RuntimeDyldELF::relocationNeedsGot(const RelocationRef &R) const {
   unsigned RelTy = R.getType();
   if (Arch == Triple::aarch64 || Arch == Triple::aarch64_be)
     return RelTy == ELF::R_AARCH64_ADR_GOT_PAGE ||
            RelTy == ELF::R_AARCH64_LD64_GOT_LO12_NC;
 
   if (Arch == Triple::x86_64)
     return RelTy == ELF::R_X86_64_GOTPCREL ||

llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldImpl.h

unsigned computeSectionStubBufSize(const ObjectFile &Obj,
                                      const SectionRef &Section);
 
   // Implementation of the generic part of the loadObject algorithm.
   Expected<ObjSectionToIDMap> loadObjectImpl(const object::ObjectFile &Obj);
 
   // Return size of Global Offset Table (GOT) entry
   virtual size_t getGOTEntrySize() { return 0; }
 
+  // Hook for the subclasses to do further processing when a symbol is added to
+  // the global symbol table. This function may modify the symbol table entry.
+  virtual void processNewSymbol(const SymbolRef &ObjSymbol, SymbolTableEntry& Entry) {}
+
   // Return true if the relocation R may require allocating a GOT entry.
   virtual bool relocationNeedsGot(const RelocationRef &R) const {
     return false;
   }
 
   // Return true if the relocation R may require allocating a stub.
   virtual bool relocationNeedsStub(const RelocationRef &R) const {
     return true;    // Conservative answer

llvm/test/ExecutionEngine/RuntimeDyld/X86/ELF_STT_GNU_IFUNC.s

+# RUN: rm -rf %t && mkdir -p %t
+# RUN: split-file %s %t
+# RUN: llvm-mc -triple=x86_64-unknown-linux-gnu -filetype=obj -o %t/test_runner.o %t/test_runner.s
+# RUN: llvm-mc -triple=x86_64-unknown-linux-gnu -filetype=obj -o %t/func_defs.o %t/func_defs.s
+# RUN: llvm-rtdyld -triple=x86_64-unknown-linux-gnu -verify -check=%s %t/test_runner.o %t/func_defs.o
+# RUN: llvm-rtdyld -triple=x86_64-unknown-linux-gnu -execute %t/test_runner.o %t/func_defs.o
+
+#--- test_runner.s
+
+# The _main function of this file contains calls to the two external functions
+# "indirect_func" and "normal_func" that are not yet defined. They are called via
+# the PLT to simulate how a compiler would emit a call to an external function.
+# Eventually, indirect_func will resolve to a STT_GNU_IFUNC and normal_func to a
+# regular function. We include calls to both types of functions in this test to
+# test that both types of functions are executed correctly when their types are
+# not known initially.
+# It also contains a call to a locally defined indirect function. As RuntimeDyld
+# treats local functions a bit differently than external functions, we also test
+# that.
+# Verify that the functions return the excpeted value. If the external indirect
+# function call fails, this returns the error code 1. If the external normal
+# function call fails, it's the error code 2. If the call to the locally
+# defined indirect function fails, return the error code 3.
+
+local_real_func:
+  mov $0x56, %eax
+  ret
+
+local_indirect_func_resolver:
+  lea local_real_func(%rip), %rax
+  ret
+
+  .type local_indirect_func, @gnu_indirect_function
+  .set local_indirect_func, local_indirect_func_resolver
+
+  .global _main
+_main:
+  call indirect_func@plt
+  cmp $0x12, %eax
+  je 1f
+  mov $1, %eax
+  ret
+1:
+
+  call normal_func@plt
+  cmp $0x34, %eax
+  je 1f
+  mov $2, %eax
+  ret
+1:
+
+  call local_indirect_func@plt
+  cmp $0x56, %eax
+  je 1f
+  mov $3, %eax
+  ret
+1:
+
+  xor %eax, %eax
+  ret
+
+# Test that the indirect functions have the same addresses in both calls.
+# rtdyld-check: decode_operand(test_indirect_func_address_1, 4) + next_pc(test_indirect_func_address_1) = decode_operand(test_indirect_func_address_2, 4) + next_pc(test_indirect_func_address_2)
+test_indirect_func_address_1:
+  lea indirect_func(%rip), %rax
+
+test_indirect_func_address_2:
+  lea indirect_func(%rip), %rax
+
+# rtdyld-check: decode_operand(test_local_indirect_func_address_1, 4) + next_pc(test_indirect_func_address_1) = decode_operand(test_local_indirect_func_address_2, 4) + next_pc(test_indirect_func_address_2)
+test_local_indirect_func_address_1:
+  lea local_indirect_func(%rip), %rax
+
+test_local_indirect_func_address_2:
+  lea local_indirect_func(%rip), %rax
+
+#--- func_defs.s
+
+# This file contains the external functions that are called above. The type of
+# the indirect function is set to @gnu_indirect_function and its value is set
+# to the value of ifunc_resolver. This is what gcc emits when using
+# __attribute__((ifunc("ifunc_resolver"))) in C. The resolver function just
+# returns the address of the real function "real_func".
+# To test that everyting works correctly, the indirect function returns 0x12
+# and the direct function returns 0x23. This is verified in the _main function
+# above.
+
+real_func:
+  mov $0x12, %eax
+  ret
+
+ifunc_resolver:
+  lea real_func(%rip), %rax
+  ret
+
+  .global indirect_func
+  .type indirect_func, @gnu_indirect_function
+  .set indirect_func, ifunc_resolver
+
+  .global normal_func
+normal_func:
+  mov $0x34, %eax
+  ret
+
+# Test that the address of the indirect function is equal even when it is
+# defined in another object file.
+# rtdyld-check: decode_operand(test_indirect_func_address_1, 4) + next_pc(test_indirect_func_address_1) = decode_operand(test_indirect_func_address_3, 4) + next_pc(test_indirect_func_address_3)
+test_indirect_func_address_3:
+  lea indirect_func(%rip), %rax