[RuntimeDyld] Handle endiannes differences between the host and target while
[oota-llvm.git] / lib / ExecutionEngine / RuntimeDyld / RuntimeDyldMachO.cpp
index 999e8f32617806302c4a85c6bd368faa92425237..e07c5dfeee5a674673f71e0dd943e552cbfe830e 100644 (file)
@@ -14,6 +14,8 @@
 #include "RuntimeDyldMachO.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/StringRef.h"
+#include "ObjectImageCommon.h"
+#include "JITRegistrar.h"
 using namespace llvm;
 using namespace llvm::object;
 
@@ -21,8 +23,144 @@ using namespace llvm::object;
 
 namespace llvm {
 
+class MachOObjectImage : public ObjectImageCommon {
+private:
+  typedef SmallVector<uint64_t, 1> SectionAddrList;
+  SectionAddrList OldSectionAddrList;
+
+protected:
+  bool is64;
+  bool Registered;
+
+private:
+  void initOldAddress() {
+    MachOObjectFile *objf = static_cast<MachOObjectFile *>(ObjFile.get());
+    // Unfortunately we need to do this, since there's information encoded
+    // in the original addr of the section that we could not otherwise
+    // recover. The reason for this is that symbols do not actually store
+    // their file offset, but only their vmaddr. This means that in order
+    // to locate the symbol correctly in the object file, we need to know
+    // where the original start of the section was (including any padding,
+    // etc).
+    for (section_iterator i = objf->section_begin(), e = objf->section_end();
+         i != e; ++i) {
+      uint64_t Addr;
+      i->getAddress(Addr);
+      OldSectionAddrList[i->getRawDataRefImpl().d.a] = Addr;
+    }
+  }
+
+public:
+  MachOObjectImage(ObjectBuffer *Input, bool is64)
+      : ObjectImageCommon(Input),
+        OldSectionAddrList(ObjFile->section_end()->getRawDataRefImpl().d.a, 0),
+        is64(is64), Registered(false) {
+    initOldAddress();
+  }
+
+  MachOObjectImage(std::unique_ptr<object::ObjectFile> Input, bool is64)
+      : ObjectImageCommon(std::move(Input)),
+        OldSectionAddrList(ObjFile->section_end()->getRawDataRefImpl().d.a, 0),
+        is64(is64), Registered(false) {
+    initOldAddress();
+  }
+
+  virtual ~MachOObjectImage() {
+    if (Registered)
+      deregisterWithDebugger();
+  }
+
+  // Subclasses can override these methods to update the image with loaded
+  // addresses for sections and common symbols
+  virtual void updateSectionAddress(const SectionRef &Sec, uint64_t Addr) {
+    MachOObjectFile *objf = static_cast<MachOObjectFile *>(ObjFile.get());
+    char *data =
+        const_cast<char *>(objf->getSectionPointer(Sec.getRawDataRefImpl()));
+
+    uint64_t oldAddr = OldSectionAddrList[Sec.getRawDataRefImpl().d.a];
+
+    if (is64) {
+      ((MachO::section_64 *)data)->addr = Addr;
+    } else {
+      ((MachO::section *)data)->addr = Addr;
+    }
+
+    for (symbol_iterator i = objf->symbol_begin(), e = objf->symbol_end();
+         i != e; ++i) {
+      section_iterator symSec(objf->section_end());
+      (*i).getSection(symSec);
+      if (*symSec == Sec) {
+        uint64_t symAddr;
+        (*i).getAddress(symAddr);
+        updateSymbolAddress(*i, symAddr + Addr - oldAddr);
+      }
+    }
+  }
+
+  uint64_t getOldSectionAddr(const SectionRef &Sec) const {
+    return OldSectionAddrList[Sec.getRawDataRefImpl().d.a];
+  }
+
+  virtual void updateSymbolAddress(const SymbolRef &Sym, uint64_t Addr) {
+    char *data = const_cast<char *>(
+        reinterpret_cast<const char *>(Sym.getRawDataRefImpl().p));
+    if (is64)
+      ((MachO::nlist_64 *)data)->n_value = Addr;
+    else
+      ((MachO::nlist *)data)->n_value = Addr;
+  }
+
+  virtual void registerWithDebugger() {
+    JITRegistrar::getGDBRegistrar().registerObject(*Buffer);
+    Registered = true;
+  }
+
+  virtual void deregisterWithDebugger() {
+    JITRegistrar::getGDBRegistrar().deregisterObject(*Buffer);
+  }
+};
+
+static uint32_t readMachOMagic(const char *InputBuffer, unsigned BufferSize) {
+  if (BufferSize < 4)
+    return 0;
+  StringRef Magic(InputBuffer, 4);
+  if (Magic == "\xFE\xED\xFA\xCE" || Magic == "\xCE\xFA\xED\xFE")
+    return 0xFEEDFACE;
+  else if (Magic == "\xFE\xED\xFA\xCF" || Magic == "\xCF\xFA\xED\xFE")
+    return 0xFEEDFACF;
+  // else
+  return 0;
+}
+
+ObjectImage *RuntimeDyldMachO::createObjectImage(ObjectBuffer *Buffer) {
+  uint32_t magic = readMachOMagic(Buffer->getBufferStart(),
+                                  Buffer->getBufferSize());
+  bool is64 = (magic == MachO::MH_MAGIC_64);
+  assert((magic == MachO::MH_MAGIC_64 || magic == MachO::MH_MAGIC) &&
+         "Unrecognized Macho Magic");
+  return new MachOObjectImage(Buffer, is64);
+}
+
+ObjectImage *RuntimeDyldMachO::createObjectImageFromFile(
+    std::unique_ptr<object::ObjectFile> ObjFile) {
+  if (!ObjFile)
+    return nullptr;
+
+  MemoryBuffer *Buffer =
+      MemoryBuffer::getMemBuffer(ObjFile->getData(), "", false);
+
+  uint32_t magic = readMachOMagic(Buffer->getBufferStart(),
+                                  Buffer->getBufferSize());
+  bool is64 = (magic == MachO::MH_MAGIC_64);
+  assert((magic == MachO::MH_MAGIC_64 || magic == MachO::MH_MAGIC) &&
+         "Unrecognized Macho Magic");
+  return new MachOObjectImage(std::move(ObjFile), is64);
+}
+
 static unsigned char *processFDE(unsigned char *P, intptr_t DeltaForText,
                                  intptr_t DeltaForEH) {
+  DEBUG(dbgs() << "Processing FDE: Delta for text: " << DeltaForText
+               << ", Delta for EH: " << DeltaForEH << "\n");
   uint32_t Length = *((uint32_t *)P);
   P += 4;
   unsigned char *Ret = P + Length;
@@ -67,7 +205,7 @@ void RuntimeDyldMachO::registerEHFrames() {
       continue;
     SectionEntry *Text = &Sections[SectionInfo.TextSID];
     SectionEntry *EHFrame = &Sections[SectionInfo.EHFrameSID];
-    SectionEntry *ExceptTab = NULL;
+    SectionEntry *ExceptTab = nullptr;
     if (SectionInfo.ExceptTabSID != RTDYLD_INVALID_SECTION_ID)
       ExceptTab = &Sections[SectionInfo.ExceptTabSID];
 
@@ -88,7 +226,8 @@ void RuntimeDyldMachO::registerEHFrames() {
   UnregisteredEHFrameSections.clear();
 }
 
-void RuntimeDyldMachO::finalizeLoad(ObjSectionToIDMap &SectionMap) {
+void RuntimeDyldMachO::finalizeLoad(ObjectImage &ObjImg,
+                                    ObjSectionToIDMap &SectionMap) {
   unsigned EHFrameSID = RTDYLD_INVALID_SECTION_ID;
   unsigned TextSID = RTDYLD_INVALID_SECTION_ID;
   unsigned ExceptTabSID = RTDYLD_INVALID_SECTION_ID;
@@ -103,6 +242,12 @@ void RuntimeDyldMachO::finalizeLoad(ObjSectionToIDMap &SectionMap) {
       TextSID = i->second;
     else if (Name == "__gcc_except_tab")
       ExceptTabSID = i->second;
+    else if (Name == "__jump_table")
+      populateJumpTable(cast<MachOObjectFile>(*ObjImg.getObjectFile()),
+                        Section, i->second);
+    else if (Name == "__pointers")
+      populatePointersSection(cast<MachOObjectFile>(*ObjImg.getObjectFile()),
+                              Section, i->second);
   }
   UnregisteredEHFrameSections.push_back(
       EHFrameRelatedSections(EHFrameSID, TextSID, ExceptTabSID));
@@ -130,91 +275,87 @@ void RuntimeDyldMachO::finalizeLoad(ObjSectionToIDMap &SectionMap) {
 // symbol in the target address space.
 void RuntimeDyldMachO::resolveRelocation(const RelocationEntry &RE,
                                          uint64_t Value) {
-  const SectionEntry &Section = Sections[RE.SectionID];
-  return resolveRelocation(Section, RE.Offset, Value, RE.RelType, RE.Addend,
-                           RE.IsPCRel, RE.Size);
-}
-
-void RuntimeDyldMachO::resolveRelocation(const SectionEntry &Section,
-                                         uint64_t Offset, uint64_t Value,
-                                         uint32_t Type, int64_t Addend,
-                                         bool isPCRel, unsigned LogSize) {
-  uint8_t *LocalAddress = Section.Address + Offset;
-  uint64_t FinalAddress = Section.LoadAddress + Offset;
-  unsigned MachoType = Type;
-  unsigned Size = 1 << LogSize;
-
-  DEBUG(dbgs() << "resolveRelocation LocalAddress: "
-               << format("%p", LocalAddress)
-               << " FinalAddress: " << format("%p", FinalAddress)
-               << " Value: " << format("%p", Value) << " Addend: " << Addend
-               << " isPCRel: " << isPCRel << " MachoType: " << MachoType
-               << " Size: " << Size << "\n");
+  DEBUG (
+    const SectionEntry &Section = Sections[RE.SectionID];
+    uint8_t* LocalAddress = Section.Address + RE.Offset;
+    uint64_t FinalAddress = Section.LoadAddress + RE.Offset;
+
+    dbgs() << "resolveRelocation Section: " << RE.SectionID
+           << " LocalAddress: " << format("%p", LocalAddress)
+           << " FinalAddress: " << format("%p", FinalAddress)
+           << " Value: " << format("%p", Value)
+           << " Addend: " << RE.Addend
+           << " isPCRel: " << RE.IsPCRel
+           << " MachoType: " << RE.RelType
+           << " Size: " << (1 << RE.Size) << "\n";
+  );
 
   // This just dispatches to the proper target specific routine.
   switch (Arch) {
   default:
     llvm_unreachable("Unsupported CPU type!");
   case Triple::x86_64:
-    resolveX86_64Relocation(LocalAddress, FinalAddress, (uintptr_t)Value,
-                            isPCRel, MachoType, Size, Addend);
+    resolveX86_64Relocation(RE, Value);
     break;
   case Triple::x86:
-    resolveI386Relocation(LocalAddress, FinalAddress, (uintptr_t)Value, isPCRel,
-                          MachoType, Size, Addend);
+    resolveI386Relocation(RE, Value);
     break;
   case Triple::arm: // Fall through.
   case Triple::thumb:
-    resolveARMRelocation(LocalAddress, FinalAddress, (uintptr_t)Value, isPCRel,
-                         MachoType, Size, Addend);
+    resolveARMRelocation(RE, Value);
     break;
+  case Triple::aarch64:
   case Triple::arm64:
-    resolveARM64Relocation(LocalAddress, FinalAddress, (uintptr_t)Value,
-                           isPCRel, MachoType, Size, Addend);
+    resolveAArch64Relocation(RE, Value);
     break;
   }
 }
 
-bool RuntimeDyldMachO::resolveI386Relocation(uint8_t *LocalAddress,
-                                             uint64_t FinalAddress,
-                                             uint64_t Value, bool isPCRel,
-                                             unsigned Type, unsigned Size,
-                                             int64_t Addend) {
-  if (isPCRel)
-    Value -= FinalAddress + 4; // see resolveX86_64Relocation
+bool RuntimeDyldMachO::resolveI386Relocation(const RelocationEntry &RE,
+                                             uint64_t Value) {
+  const SectionEntry &Section = Sections[RE.SectionID];
+  uint8_t* LocalAddress = Section.Address + RE.Offset;
 
-  switch (Type) {
-  default:
-    llvm_unreachable("Invalid relocation type!");
-  case MachO::GENERIC_RELOC_VANILLA: {
-    uint8_t *p = LocalAddress;
-    uint64_t ValueToWrite = Value + Addend;
-    for (unsigned i = 0; i < Size; ++i) {
-      *p++ = (uint8_t)(ValueToWrite & 0xff);
-      ValueToWrite >>= 8;
-    }
-    return false;
+  if (RE.IsPCRel) {
+    uint64_t FinalAddress = Section.LoadAddress + RE.Offset;
+    Value -= FinalAddress + 4; // see MachOX86_64::resolveRelocation.
   }
-  case MachO::GENERIC_RELOC_SECTDIFF:
-  case MachO::GENERIC_RELOC_LOCAL_SECTDIFF:
-  case MachO::GENERIC_RELOC_PB_LA_PTR:
-    return Error("Relocation type not implemented yet!");
+
+  switch (RE.RelType) {
+    default:
+      llvm_unreachable("Invalid relocation type!");
+    case MachO::GENERIC_RELOC_VANILLA:
+      return applyRelocationValue(LocalAddress, Value + RE.Addend,
+                                  1 << RE.Size);
+    case MachO::GENERIC_RELOC_SECTDIFF:
+    case MachO::GENERIC_RELOC_LOCAL_SECTDIFF: {
+      uint64_t SectionABase = Sections[RE.Sections.SectionA].LoadAddress;
+      uint64_t SectionBBase = Sections[RE.Sections.SectionB].LoadAddress;
+      assert((Value == SectionABase || Value == SectionBBase) &&
+             "Unexpected SECTDIFF relocation value.");
+      Value = SectionABase - SectionBBase + RE.Addend;
+      return applyRelocationValue(LocalAddress, Value, 1 << RE.Size);
+    }
+    case MachO::GENERIC_RELOC_PB_LA_PTR:
+      return Error("Relocation type not implemented yet!");
   }
 }
 
-bool RuntimeDyldMachO::resolveX86_64Relocation(uint8_t *LocalAddress,
-                                               uint64_t FinalAddress,
-                                               uint64_t Value, bool isPCRel,
-                                               unsigned Type, unsigned Size,
-                                               int64_t Addend) {
+bool RuntimeDyldMachO::resolveX86_64Relocation(const RelocationEntry &RE,
+                                               uint64_t Value) {
+  const SectionEntry &Section = Sections[RE.SectionID];
+  uint8_t* LocalAddress = Section.Address + RE.Offset;
+
   // If the relocation is PC-relative, the value to be encoded is the
   // pointer difference.
-  if (isPCRel)
+  if (RE.IsPCRel) {
     // FIXME: It seems this value needs to be adjusted by 4 for an effective PC
     // address. Is that expected? Only for branches, perhaps?
-    Value -= FinalAddress + 4;
+    uint64_t FinalAddress = Section.LoadAddress + RE.Offset;
+    Value -= FinalAddress + 4; // see MachOX86_64::resolveRelocation.
+  }
 
-  switch (Type) {
+  switch (RE.RelType) {
   default:
     llvm_unreachable("Invalid relocation type!");
   case MachO::X86_64_RELOC_SIGNED_1:
@@ -222,17 +363,8 @@ bool RuntimeDyldMachO::resolveX86_64Relocation(uint8_t *LocalAddress,
   case MachO::X86_64_RELOC_SIGNED_4:
   case MachO::X86_64_RELOC_SIGNED:
   case MachO::X86_64_RELOC_UNSIGNED:
-  case MachO::X86_64_RELOC_BRANCH: {
-    Value += Addend;
-    // Mask in the target value a byte at a time (we don't have an alignment
-    // guarantee for the target address, so this is safest).
-    uint8_t *p = (uint8_t *)LocalAddress;
-    for (unsigned i = 0; i < Size; ++i) {
-      *p++ = (uint8_t)Value;
-      Value >>= 8;
-    }
-    return false;
-  }
+  case MachO::X86_64_RELOC_BRANCH:
+    return applyRelocationValue(LocalAddress, Value + RE.Addend, 1 << RE.Size);
   case MachO::X86_64_RELOC_GOT_LOAD:
   case MachO::X86_64_RELOC_GOT:
   case MachO::X86_64_RELOC_SUBTRACTOR:
@@ -241,14 +373,15 @@ bool RuntimeDyldMachO::resolveX86_64Relocation(uint8_t *LocalAddress,
   }
 }
 
-bool RuntimeDyldMachO::resolveARMRelocation(uint8_t *LocalAddress,
-                                            uint64_t FinalAddress,
-                                            uint64_t Value, bool isPCRel,
-                                            unsigned Type, unsigned Size,
-                                            int64_t Addend) {
+bool RuntimeDyldMachO::resolveARMRelocation(const RelocationEntry &RE,
+                                            uint64_t Value) {
+  const SectionEntry &Section = Sections[RE.SectionID];
+  uint8_t* LocalAddress = Section.Address + RE.Offset;
+
   // If the relocation is PC-relative, the value to be encoded is the
   // pointer difference.
-  if (isPCRel) {
+  if (RE.IsPCRel) {
+    uint64_t FinalAddress = Section.LoadAddress + RE.Offset;
     Value -= FinalAddress;
     // ARM PCRel relocations have an effective-PC offset of two instructions
     // (four bytes in Thumb mode, 8 bytes in ARM mode).
@@ -256,19 +389,11 @@ bool RuntimeDyldMachO::resolveARMRelocation(uint8_t *LocalAddress,
     Value -= 8;
   }
 
-  switch (Type) {
+  switch (RE.RelType) {
   default:
     llvm_unreachable("Invalid relocation type!");
-  case MachO::ARM_RELOC_VANILLA: {
-    // Mask in the target value a byte at a time (we don't have an alignment
-    // guarantee for the target address, so this is safest).
-    uint8_t *p = (uint8_t *)LocalAddress;
-    for (unsigned i = 0; i < Size; ++i) {
-      *p++ = (uint8_t)Value;
-      Value >>= 8;
-    }
-    break;
-  }
+  case MachO::ARM_RELOC_VANILLA:
+    return applyRelocationValue(LocalAddress, Value, 1 << RE.Size);
   case MachO::ARM_RELOC_BR24: {
     // Mask the value into the target address. We know instructions are
     // 32-bit aligned, so we can do it all at once.
@@ -276,13 +401,16 @@ bool RuntimeDyldMachO::resolveARMRelocation(uint8_t *LocalAddress,
     // The low two bits of the value are not encoded.
     Value >>= 2;
     // Mask the value to 24 bits.
-    Value &= 0xffffff;
+    uint64_t FinalValue = Value & 0xffffff;
+    // Check for overflow.
+    if (Value != FinalValue)
+      return Error("ARM BR24 relocation out of range.");
     // FIXME: If the destination is a Thumb function (and the instruction
     // is a non-predicated BL instruction), we need to change it to a BLX
     // instruction instead.
 
     // Insert the value into the instruction.
-    *p = (*p & ~0xffffff) | Value;
+    *p = (*p & ~0xffffff) | FinalValue;
     break;
   }
   case MachO::ARM_THUMB_RELOC_BR22:
@@ -298,89 +426,405 @@ bool RuntimeDyldMachO::resolveARMRelocation(uint8_t *LocalAddress,
   return false;
 }
 
-bool RuntimeDyldMachO::resolveARM64Relocation(uint8_t *LocalAddress,
-                                              uint64_t FinalAddress,
-                                              uint64_t Value, bool isPCRel,
-                                              unsigned Type, unsigned Size,
-                                              int64_t Addend) {
-  // If the relocation is PC-relative, the value to be encoded is the
-  // pointer difference.
-  if (isPCRel)
-    Value -= FinalAddress;
+bool RuntimeDyldMachO::resolveAArch64Relocation(const RelocationEntry &RE,
+                                                uint64_t Value) {
+  const SectionEntry &Section = Sections[RE.SectionID];
+  uint8_t* LocalAddress = Section.Address + RE.Offset;
 
-  switch (Type) {
+  switch (RE.RelType) {
   default:
     llvm_unreachable("Invalid relocation type!");
   case MachO::ARM64_RELOC_UNSIGNED: {
+    assert(!RE.IsPCRel && "PCRel and ARM64_RELOC_UNSIGNED not supported");
     // Mask in the target value a byte at a time (we don't have an alignment
     // guarantee for the target address, so this is safest).
-    uint8_t *p = (uint8_t *)LocalAddress;
-    for (unsigned i = 0; i < Size; ++i) {
-      *p++ = (uint8_t)Value;
-      Value >>= 8;
-    }
+    if (RE.Size < 2)
+      llvm_unreachable("Invalid size for ARM64_RELOC_UNSIGNED");
+
+    applyRelocationValue(LocalAddress, Value + RE.Addend, 1 << RE.Size);
     break;
   }
   case MachO::ARM64_RELOC_BRANCH26: {
+    assert(RE.IsPCRel && "not PCRel and ARM64_RELOC_BRANCH26 not supported");
     // Mask the value into the target address. We know instructions are
     // 32-bit aligned, so we can do it all at once.
-    uint32_t *p = (uint32_t *)LocalAddress;
-    // The low two bits of the value are not encoded.
-    Value >>= 2;
-    // Mask the value to 26 bits.
-    Value &= 0x3ffffff;
+    uint32_t *p = (uint32_t*)LocalAddress;
+    // Check if the addend is encoded in the instruction.
+    uint32_t EncodedAddend = *p & 0x03FFFFFF;
+    if (EncodedAddend != 0 ) {
+      if (RE.Addend == 0)
+        llvm_unreachable("branch26 instruction has embedded addend.");
+      else
+        llvm_unreachable("branch26 instruction has embedded addend and" \
+                         "ARM64_RELOC_ADDEND.");
+    }
+    // Check if branch is in range.
+    uint64_t FinalAddress = Section.LoadAddress + RE.Offset;
+    uint64_t PCRelVal = Value - FinalAddress + RE.Addend;
+    assert(isInt<26>(PCRelVal) && "Branch target out of range!");
     // Insert the value into the instruction.
-    *p = (*p & ~0x3ffffff) | Value;
+    *p = (*p & 0xFC000000) | ((uint32_t)(PCRelVal >> 2) & 0x03FFFFFF);
     break;
   }
-  case MachO::ARM64_RELOC_SUBTRACTOR:
-  case MachO::ARM64_RELOC_PAGE21:
-  case MachO::ARM64_RELOC_PAGEOFF12:
   case MachO::ARM64_RELOC_GOT_LOAD_PAGE21:
+  case MachO::ARM64_RELOC_PAGE21: {
+    assert(RE.IsPCRel && "not PCRel and ARM64_RELOC_PAGE21 not supported");
+    // Mask the value into the target address. We know instructions are
+    // 32-bit aligned, so we can do it all at once.
+    uint32_t *p = (uint32_t*)LocalAddress;
+    // Check if the addend is encoded in the instruction.
+    uint32_t EncodedAddend = ((*p & 0x60000000) >> 29) |
+                             ((*p & 0x01FFFFE0) >> 3);
+    if (EncodedAddend != 0) {
+      if (RE.Addend == 0)
+        llvm_unreachable("adrp instruction has embedded addend.");
+      else
+        llvm_unreachable("adrp instruction has embedded addend and" \
+                         "ARM64_RELOC_ADDEND.");
+    }
+    // Adjust for PC-relative relocation and offset.
+    uint64_t FinalAddress = Section.LoadAddress + RE.Offset;
+    uint64_t PCRelVal = ((Value + RE.Addend) & (-4096)) -
+                         (FinalAddress & (-4096));
+    // Check that the value fits into 21 bits (+ 12 lower bits).
+    assert(isInt<33>(PCRelVal) && "Invalid page reloc value!");
+    // Insert the value into the instruction.
+    uint32_t ImmLoValue = (uint32_t)(PCRelVal << 17) & 0x60000000;
+    uint32_t ImmHiValue = (uint32_t)(PCRelVal >>  9) & 0x00FFFFE0;
+    *p = (*p & 0x9F00001F) | ImmHiValue | ImmLoValue;
+    break;
+  }
   case MachO::ARM64_RELOC_GOT_LOAD_PAGEOFF12:
+  case MachO::ARM64_RELOC_PAGEOFF12: {
+    assert(!RE.IsPCRel && "PCRel and ARM64_RELOC_PAGEOFF21 not supported");
+    // Mask the value into the target address. We know instructions are
+    // 32-bit aligned, so we can do it all at once.
+    uint32_t *p = (uint32_t*)LocalAddress;
+    // Check if the addend is encoded in the instruction.
+    uint32_t EncodedAddend = *p & 0x003FFC00;
+    if (EncodedAddend != 0) {
+      if (RE.Addend == 0)
+        llvm_unreachable("adrp instruction has embedded addend.");
+      else
+        llvm_unreachable("adrp instruction has embedded addend and" \
+                         "ARM64_RELOC_ADDEND.");
+    }
+    // Add the offset from the symbol.
+    Value += RE.Addend;
+    // Mask out the page address and only use the lower 12 bits.
+    Value &= 0xFFF;
+    // Check which instruction we are updating to obtain the implicit shift
+    // factor from LDR/STR instructions.
+    if (*p & 0x08000000) {
+      uint32_t ImplicitShift = ((*p >> 30) & 0x3);
+      switch (ImplicitShift) {
+      case 0:
+        // Check if this a vector op.
+        if ((*p & 0x04800000) == 0x04800000) {
+          ImplicitShift = 4;
+          assert(((Value & 0xF) == 0) &&
+                 "128-bit LDR/STR not 16-byte aligned.");
+        }
+        break;
+      case 1:
+        assert(((Value & 0x1) == 0) && "16-bit LDR/STR not 2-byte aligned.");
+      case 2:
+        assert(((Value & 0x3) == 0) && "32-bit LDR/STR not 4-byte aligned.");
+      case 3:
+        assert(((Value & 0x7) == 0) && "64-bit LDR/STR not 8-byte aligned.");
+      }
+      // Compensate for implicit shift.
+      Value >>= ImplicitShift;
+    }
+    // Insert the value into the instruction.
+    *p = (*p & 0xFFC003FF) | ((uint32_t)(Value << 10) & 0x003FFC00);
+    break;
+  }
+  case MachO::ARM64_RELOC_SUBTRACTOR:
   case MachO::ARM64_RELOC_POINTER_TO_GOT:
   case MachO::ARM64_RELOC_TLVP_LOAD_PAGE21:
   case MachO::ARM64_RELOC_TLVP_LOAD_PAGEOFF12:
-  case MachO::ARM64_RELOC_ADDEND:
+    llvm_unreachable("Relocation type not implemented yet!");
     return Error("Relocation type not implemented yet!");
+  case MachO::ARM64_RELOC_ADDEND:
+    llvm_unreachable("ARM64_RELOC_ADDEND should have been handeled by " \
+                     "processRelocationRef!");
   }
   return false;
 }
 
+void RuntimeDyldMachO::populateJumpTable(MachOObjectFile &Obj,
+                                         const SectionRef &JTSection,
+                                         unsigned JTSectionID) {
+  assert(!Obj.is64Bit() &&
+         "__jump_table section not supported in 64-bit MachO.");
+
+  MachO::dysymtab_command DySymTabCmd = Obj.getDysymtabLoadCommand();
+  MachO::section Sec32 = Obj.getSection(JTSection.getRawDataRefImpl());
+  uint32_t JTSectionSize = Sec32.size;
+  unsigned FirstIndirectSymbol = Sec32.reserved1;
+  unsigned JTEntrySize = Sec32.reserved2;
+  unsigned NumJTEntries = JTSectionSize / JTEntrySize;
+  uint8_t* JTSectionAddr = getSectionAddress(JTSectionID);
+  unsigned JTEntryOffset = 0;
+
+  assert((JTSectionSize % JTEntrySize) == 0 &&
+         "Jump-table section does not contain a whole number of stubs?");
+
+  for (unsigned i = 0; i < NumJTEntries; ++i) {
+    unsigned SymbolIndex =
+      Obj.getIndirectSymbolTableEntry(DySymTabCmd, FirstIndirectSymbol + i);
+    symbol_iterator SI = Obj.getSymbolByIndex(SymbolIndex);
+    StringRef IndirectSymbolName;
+    SI->getName(IndirectSymbolName);
+    uint8_t* JTEntryAddr = JTSectionAddr + JTEntryOffset;
+    createStubFunction(JTEntryAddr);
+    RelocationEntry RE(JTSectionID, JTEntryOffset + 1,
+                       MachO::GENERIC_RELOC_VANILLA, 0, true, 2);
+    addRelocationForSymbol(RE, IndirectSymbolName);
+    JTEntryOffset += JTEntrySize;
+  }
+}
+
+void RuntimeDyldMachO::populatePointersSection(MachOObjectFile &Obj,
+                                               const SectionRef &PTSection,
+                                               unsigned PTSectionID) {
+  assert(!Obj.is64Bit() &&
+         "__pointers section not supported in 64-bit MachO.");
+
+  MachO::dysymtab_command DySymTabCmd = Obj.getDysymtabLoadCommand();
+  MachO::section Sec32 = Obj.getSection(PTSection.getRawDataRefImpl());
+  uint32_t PTSectionSize = Sec32.size;
+  unsigned FirstIndirectSymbol = Sec32.reserved1;
+  const unsigned PTEntrySize = 4;
+  unsigned NumPTEntries = PTSectionSize / PTEntrySize;
+  unsigned PTEntryOffset = 0;
+
+  assert((PTSectionSize % PTEntrySize) == 0 &&
+         "Pointers section does not contain a whole number of stubs?");
+
+  DEBUG(dbgs() << "Populating __pointers, Section ID " << PTSectionID
+               << ", " << NumPTEntries << " entries, "
+               << PTEntrySize << " bytes each:\n");
+
+  for (unsigned i = 0; i < NumPTEntries; ++i) {
+    unsigned SymbolIndex =
+      Obj.getIndirectSymbolTableEntry(DySymTabCmd, FirstIndirectSymbol + i);
+    symbol_iterator SI = Obj.getSymbolByIndex(SymbolIndex);
+    StringRef IndirectSymbolName;
+    SI->getName(IndirectSymbolName);
+    DEBUG(dbgs() << "  " << IndirectSymbolName << ": index " << SymbolIndex
+          << ", PT offset: " << PTEntryOffset << "\n");
+    RelocationEntry RE(PTSectionID, PTEntryOffset,
+                       MachO::GENERIC_RELOC_VANILLA, 0, false, 2);
+    addRelocationForSymbol(RE, IndirectSymbolName);
+    PTEntryOffset += PTEntrySize;
+  }
+}
+
+
+section_iterator getSectionByAddress(const MachOObjectFile &Obj,
+                                     uint64_t Addr) {
+  section_iterator SI = Obj.section_begin();
+  section_iterator SE = Obj.section_end();
+
+  for (; SI != SE; ++SI) {
+    uint64_t SAddr, SSize;
+    SI->getAddress(SAddr);
+    SI->getSize(SSize);
+    if ((Addr >= SAddr) && (Addr < SAddr + SSize))
+      return SI;
+  }
+
+  return SE;
+}
+
+relocation_iterator RuntimeDyldMachO::processSECTDIFFRelocation(
+                                            unsigned SectionID,
+                                            relocation_iterator RelI,
+                                            ObjectImage &Obj,
+                                            ObjSectionToIDMap &ObjSectionToID) {
+  const MachOObjectFile *MachO =
+    static_cast<const MachOObjectFile*>(Obj.getObjectFile());
+  MachO::any_relocation_info RE =
+    MachO->getRelocation(RelI->getRawDataRefImpl());
+
+  SectionEntry &Section = Sections[SectionID];
+  uint32_t RelocType = MachO->getAnyRelocationType(RE);
+  bool IsPCRel = MachO->getAnyRelocationPCRel(RE);
+  unsigned Size = MachO->getAnyRelocationLength(RE);
+  uint64_t Offset;
+  RelI->getOffset(Offset);
+  uint8_t *LocalAddress = Section.Address + Offset;
+  unsigned NumBytes = 1 << Size;
+  int64_t Addend = 0;
+  memcpy(&Addend, LocalAddress, NumBytes);
+
+  ++RelI;
+  MachO::any_relocation_info RE2 =
+    MachO->getRelocation(RelI->getRawDataRefImpl());
+
+  uint32_t AddrA = MachO->getScatteredRelocationValue(RE);
+  section_iterator SAI = getSectionByAddress(*MachO, AddrA);
+  assert(SAI != MachO->section_end() && "Can't find section for address A");
+  uint64_t SectionABase;
+  SAI->getAddress(SectionABase);
+  uint64_t SectionAOffset = AddrA - SectionABase;
+  SectionRef SectionA = *SAI;
+  bool IsCode;
+  SectionA.isText(IsCode);
+  uint32_t SectionAID = findOrEmitSection(Obj, SectionA, IsCode,
+                                          ObjSectionToID);
+
+  uint32_t AddrB = MachO->getScatteredRelocationValue(RE2);
+  section_iterator SBI = getSectionByAddress(*MachO, AddrB);
+  assert(SBI != MachO->section_end() && "Can't find section for address B");
+  uint64_t SectionBBase;
+  SBI->getAddress(SectionBBase);
+  uint64_t SectionBOffset = AddrB - SectionBBase;
+  SectionRef SectionB = *SBI;
+  uint32_t SectionBID = findOrEmitSection(Obj, SectionB, IsCode,
+                                          ObjSectionToID);
+
+  if (Addend != AddrA - AddrB)
+    Error("Unexpected SECTDIFF relocation addend.");
+
+  DEBUG(dbgs() << "Found SECTDIFF: AddrA: " << AddrA << ", AddrB: " << AddrB
+               << ", Addend: " << Addend << ", SectionA ID: "
+               << SectionAID << ", SectionAOffset: " << SectionAOffset
+               << ", SectionB ID: " << SectionBID << ", SectionBOffset: "
+               << SectionBOffset << "\n");
+  RelocationEntry R(SectionID, Offset, RelocType, 0,
+                    SectionAID, SectionAOffset, SectionBID, SectionBOffset,
+                    IsPCRel, Size);
+
+  addRelocationForSection(R, SectionAID);
+  addRelocationForSection(R, SectionBID);
+
+  return ++RelI;
+}
+
+relocation_iterator RuntimeDyldMachO::processI386ScatteredVANILLA(
+                                            unsigned SectionID,
+                                            relocation_iterator RelI,
+                                            ObjectImage &Obj,
+                                            ObjSectionToIDMap &ObjSectionToID) {
+  const MachOObjectFile *MachO =
+    static_cast<const MachOObjectFile*>(Obj.getObjectFile());
+  MachO::any_relocation_info RE =
+    MachO->getRelocation(RelI->getRawDataRefImpl());
+
+  SectionEntry &Section = Sections[SectionID];
+  uint32_t RelocType = MachO->getAnyRelocationType(RE);
+  bool IsPCRel = MachO->getAnyRelocationPCRel(RE);
+  unsigned Size = MachO->getAnyRelocationLength(RE);
+  uint64_t Offset;
+  RelI->getOffset(Offset);
+  uint8_t *LocalAddress = Section.Address + Offset;
+  unsigned NumBytes = 1 << Size;
+  int64_t Addend = 0;
+  memcpy(&Addend, LocalAddress, NumBytes);
+
+  unsigned SymbolBaseAddr = MachO->getScatteredRelocationValue(RE);
+  section_iterator TargetSI = getSectionByAddress(*MachO, SymbolBaseAddr);
+  assert(TargetSI != MachO->section_end() && "Can't find section for symbol");
+  uint64_t SectionBaseAddr;
+  TargetSI->getAddress(SectionBaseAddr);
+  SectionRef TargetSection = *TargetSI;
+  bool IsCode;
+  TargetSection.isText(IsCode);
+  uint32_t TargetSectionID = findOrEmitSection(Obj, TargetSection, IsCode,
+                                               ObjSectionToID);
+
+  Addend -= SectionBaseAddr;
+  RelocationEntry R(SectionID, Offset, RelocType, Addend,
+                    IsPCRel, Size);
+
+  addRelocationForSection(R, TargetSectionID);
+
+  return ++RelI;
+}
+
 relocation_iterator RuntimeDyldMachO::processRelocationRef(
     unsigned SectionID, relocation_iterator RelI, ObjectImage &Obj,
     ObjSectionToIDMap &ObjSectionToID, const SymbolTableMap &Symbols,
     StubMap &Stubs) {
   const ObjectFile *OF = Obj.getObjectFile();
+  const MachOObjectImage &MachOObj = *static_cast<MachOObjectImage *>(&Obj);
   const MachOObjectFile *MachO = static_cast<const MachOObjectFile *>(OF);
   MachO::any_relocation_info RE =
       MachO->getRelocation(RelI->getRawDataRefImpl());
+  int64_t RelocAddendValue = 0;
+  bool HasRelocAddendValue = false;
 
   uint32_t RelType = MachO->getAnyRelocationType(RE);
+  if (Arch == Triple::arm64) {
+    // ARM64_RELOC_ADDEND provides the offset (addend) that will be used by the
+    // next relocation entry. Save the value and advance to the next relocation
+    // entry.
+    if (RelType == MachO::ARM64_RELOC_ADDEND) {
+      assert(!MachO->getPlainRelocationExternal(RE));
+      assert(!MachO->getAnyRelocationPCRel(RE));
+      assert(MachO->getAnyRelocationLength(RE) == 2);
+      uint64_t RawAddend = MachO->getPlainRelocationSymbolNum(RE);
+      // Sign-extend the 24-bit to 64-bit.
+      RelocAddendValue = RawAddend << 40;
+      RelocAddendValue >>= 40;
+      HasRelocAddendValue = true;
+
+      // Get the next entry.
+      RE = MachO->getRelocation((++RelI)->getRawDataRefImpl());
+      RelType = MachO->getAnyRelocationType(RE);
+      assert(RelType == MachO::ARM64_RELOC_BRANCH26 ||
+             RelType == MachO::ARM64_RELOC_PAGE21 ||
+             RelType == MachO::ARM64_RELOC_PAGEOFF12);
+
+    } else if (RelType == MachO::ARM64_RELOC_BRANCH26 ||
+               RelType == MachO::ARM64_RELOC_PAGE21 ||
+               RelType == MachO::ARM64_RELOC_PAGEOFF12 ||
+               RelType == MachO::ARM64_RELOC_GOT_LOAD_PAGE21 ||
+               RelType == MachO::ARM64_RELOC_GOT_LOAD_PAGEOFF12) {
+      RelocAddendValue = 0;
+      HasRelocAddendValue = true;
+    }
+  }
 
   // FIXME: Properly handle scattered relocations.
-  //        For now, optimistically skip these: they can often be ignored, as
-  //        the static linker will already have applied the relocation, and it
-  //        only needs to be reapplied if symbols move relative to one another.
-  //        Note: This will fail horribly where the relocations *do* need to be
-  //        applied, but that was already the case.
-  if (MachO->isRelocationScattered(RE))
-    return ++RelI;
+  //        Special case the couple of scattered relocations that we know how
+  //        to handle: SECTDIFF relocations, and scattered VANILLA relocations
+  //        on I386.
+  //        For all other scattered relocations, just bail out and hope for the
+  //        best, since the offsets computed by scattered relocations have often
+  //        been optimisticaly filled in by the compiler. This will fail
+  //        horribly where the relocations *do* need to be applied, but that was
+  //        already the case.
+  if (MachO->isRelocationScattered(RE)) {
+    if (RelType == MachO::GENERIC_RELOC_SECTDIFF ||
+        RelType == MachO::GENERIC_RELOC_LOCAL_SECTDIFF)
+      return processSECTDIFFRelocation(SectionID, RelI, Obj, ObjSectionToID);
+    else if (Arch == Triple::x86 && RelType == MachO::GENERIC_RELOC_VANILLA)
+      return processI386ScatteredVANILLA(SectionID, RelI, Obj, ObjSectionToID);
+    else
+      return ++RelI;
+  }
 
   RelocationValueRef Value;
   SectionEntry &Section = Sections[SectionID];
 
-  bool isExtern = MachO->getPlainRelocationExternal(RE);
+  bool IsExtern = MachO->getPlainRelocationExternal(RE);
   bool IsPCRel = MachO->getAnyRelocationPCRel(RE);
   unsigned Size = MachO->getAnyRelocationLength(RE);
   uint64_t Offset;
   RelI->getOffset(Offset);
   uint8_t *LocalAddress = Section.Address + Offset;
   unsigned NumBytes = 1 << Size;
-  uint64_t Addend = 0;
-  memcpy(&Addend, LocalAddress, NumBytes);
+  int64_t Addend = 0;
+  if (HasRelocAddendValue)
+    Addend = RelocAddendValue;
+  else
+    memcpy(&Addend, LocalAddress, NumBytes);
 
-  if (isExtern) {
+  if (IsExtern) {
     // Obtain the symbol name which is referenced in the relocation
     symbol_iterator Symbol = RelI->getSymbol();
     StringRef TargetName;
@@ -402,13 +846,24 @@ relocation_iterator RuntimeDyldMachO::processRelocationRef(
         Value.Addend = Addend;
       }
     }
+
+    // Addends for external, PC-rel relocations on i386 point back to the zero
+    // offset. Calculate the final offset from the relocation target instead.
+    // This allows us to use the same logic for both external and internal
+    // relocations in resolveI386RelocationRef.
+    if (Arch == Triple::x86 && IsPCRel) {
+      uint64_t RelocAddr = 0;
+      RelI->getAddress(RelocAddr);
+      Value.Addend += RelocAddr + 4;
+    }
+
   } else {
     SectionRef Sec = MachO->getRelocationSection(RE);
     bool IsCode = false;
     Sec.isText(IsCode);
     Value.SectionID = findOrEmitSection(Obj, Sec, IsCode, ObjSectionToID);
-    uint64_t Addr;
-    Sec.getAddress(Addr);
+    uint64_t Addr = MachOObj.getOldSectionAddr(Sec);
+    DEBUG(dbgs() << "\nAddr: " << Addr << "\nAddend: " << Addend);
     Value.Addend = Addend - Addr;
     if (IsPCRel)
       Value.Addend += Offset + NumBytes;
@@ -418,6 +873,10 @@ relocation_iterator RuntimeDyldMachO::processRelocationRef(
                                  RelType == MachO::X86_64_RELOC_GOT_LOAD)) {
     assert(IsPCRel);
     assert(Size == 2);
+
+    // FIXME: Teach the generic code above not to prematurely conflate
+    //        relocation addends and symbol offsets.
+    Value.Addend -= Addend;
     StubMap::const_iterator i = Stubs.find(Value);
     uint8_t *Addr;
     if (i != Stubs.end()) {
@@ -425,42 +884,80 @@ relocation_iterator RuntimeDyldMachO::processRelocationRef(
     } else {
       Stubs[Value] = Section.StubOffset;
       uint8_t *GOTEntry = Section.Address + Section.StubOffset;
-      RelocationEntry RE(SectionID, Section.StubOffset,
-                         MachO::X86_64_RELOC_UNSIGNED, 0, false, 3);
+      RelocationEntry GOTRE(SectionID, Section.StubOffset,
+                            MachO::X86_64_RELOC_UNSIGNED, Value.Addend, false,
+                            3);
       if (Value.SymbolName)
-        addRelocationForSymbol(RE, Value.SymbolName);
+        addRelocationForSymbol(GOTRE, Value.SymbolName);
       else
-        addRelocationForSection(RE, Value.SectionID);
+        addRelocationForSection(GOTRE, Value.SectionID);
       Section.StubOffset += 8;
       Addr = GOTEntry;
     }
-    resolveRelocation(Section, Offset, (uint64_t)Addr,
-                      MachO::X86_64_RELOC_UNSIGNED, Value.Addend, true, 2);
+    RelocationEntry TargetRE(SectionID, Offset,
+                             MachO::X86_64_RELOC_UNSIGNED, Addend, true,
+                             2);
+    resolveRelocation(TargetRE, (uint64_t)Addr);
   } else if (Arch == Triple::arm && (RelType & 0xf) == MachO::ARM_RELOC_BR24) {
     // This is an ARM branch relocation, need to use a stub function.
 
     //  Look up for existing stub.
     StubMap::const_iterator i = Stubs.find(Value);
-    if (i != Stubs.end())
-      resolveRelocation(Section, Offset, (uint64_t)Section.Address + i->second,
-                        RelType, 0, IsPCRel, Size);
-    else {
+    uint8_t *Addr;
+    if (i != Stubs.end()) {
+      Addr = Section.Address + i->second;
+    else {
       // Create a new stub function.
       Stubs[Value] = Section.StubOffset;
       uint8_t *StubTargetAddr =
           createStubFunction(Section.Address + Section.StubOffset);
-      RelocationEntry RE(SectionID, StubTargetAddr - Section.Address,
-                         MachO::GENERIC_RELOC_VANILLA, Value.Addend);
+      RelocationEntry StubRE(SectionID, StubTargetAddr - Section.Address,
+                             MachO::GENERIC_RELOC_VANILLA, Value.Addend);
       if (Value.SymbolName)
-        addRelocationForSymbol(RE, Value.SymbolName);
+        addRelocationForSymbol(StubRE, Value.SymbolName);
       else
-        addRelocationForSection(RE, Value.SectionID);
-      resolveRelocation(Section, Offset,
-                        (uint64_t)Section.Address + Section.StubOffset, RelType,
-                        0, IsPCRel, Size);
+        addRelocationForSection(StubRE, Value.SectionID);
+      Addr = Section.Address + Section.StubOffset;
       Section.StubOffset += getMaxStubSize();
     }
+    RelocationEntry TargetRE(Value.SectionID, Offset, RelType, 0, IsPCRel,
+                             Size);
+    resolveRelocation(TargetRE, (uint64_t)Addr);
+  } else if (Arch == Triple::arm64 &&
+             (RelType == MachO::ARM64_RELOC_GOT_LOAD_PAGE21 ||
+              RelType == MachO::ARM64_RELOC_GOT_LOAD_PAGEOFF12)) {
+    assert(Size == 2);
+    StubMap::const_iterator i = Stubs.find(Value);
+    uint8_t *Addr;
+    if (i != Stubs.end())
+      Addr = Section.Address + i->second;
+    else {
+      // FIXME: There must be a better way to do this then to check and fix the
+      // alignment every time!!!
+      uintptr_t BaseAddress = uintptr_t(Section.Address);
+      uintptr_t StubAlignment = getStubAlignment();
+      uintptr_t StubAddress
+        = (BaseAddress + Section.StubOffset + StubAlignment - 1) &
+          -StubAlignment;
+      unsigned StubOffset = StubAddress - BaseAddress;
+      Stubs[Value] = StubOffset;
+      assert(((StubAddress % getStubAlignment()) == 0) &&
+             "GOT entry not aligned");
+      RelocationEntry GOTRE(SectionID, StubOffset, MachO::ARM64_RELOC_UNSIGNED,
+                            Value.Addend, /*IsPCRel=*/false, /*Size=*/3);
+      if (Value.SymbolName)
+        addRelocationForSymbol(GOTRE, Value.SymbolName);
+      else
+        addRelocationForSection(GOTRE, Value.SectionID);
+      Section.StubOffset = StubOffset + getMaxStubSize();
+
+      Addr = (uint8_t *)StubAddress;
+    }
+    RelocationEntry TargetRE(SectionID, Offset, RelType, /*Addend=*/0, IsPCRel,
+                             Size);
+    resolveRelocation(TargetRE, (uint64_t)Addr);
   } else {
+
     RelocationEntry RE(SectionID, Offset, RelType, Value.Addend, IsPCRel, Size);
     if (Value.SymbolName)
       addRelocationForSymbol(RE, Value.SymbolName);
@@ -472,18 +969,9 @@ relocation_iterator RuntimeDyldMachO::processRelocationRef(
 
 bool
 RuntimeDyldMachO::isCompatibleFormat(const ObjectBuffer *InputBuffer) const {
-  if (InputBuffer->getBufferSize() < 4)
-    return false;
-  StringRef Magic(InputBuffer->getBufferStart(), 4);
-  if (Magic == "\xFE\xED\xFA\xCE")
-    return true;
-  if (Magic == "\xCE\xFA\xED\xFE")
-    return true;
-  if (Magic == "\xFE\xED\xFA\xCF")
-    return true;
-  if (Magic == "\xCF\xFA\xED\xFE")
-    return true;
-  return false;
+  uint32_t Magic = readMachOMagic(InputBuffer->getBufferStart(),
+                                  InputBuffer->getBufferSize());
+  return (Magic == 0xFEEDFACE || Magic == 0xFEEDFACF);
 }
 
 bool RuntimeDyldMachO::isCompatibleFile(const object::ObjectFile *Obj) const {