Adding a basic ELF dynamic loader and MC-JIT for ELF. Functionality is currently basic and will be enhanced with future patches.

Patch developed by Andy Kaylor and Daniel Malea. Reviewed on llvm-commits.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@148231 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 373 insertions, 5 deletions

diff --git a/lib/ExecutionEngine/RuntimeDyld/CMakeLists.txt b/lib/ExecutionEngine/RuntimeDyld/CMakeLists.txt
index 59bdfee3db..002e63cd3b 100644
--- a/lib/ExecutionEngine/RuntimeDyld/CMakeLists.txt
+++ b/lib/ExecutionEngine/RuntimeDyld/CMakeLists.txt
@@ -1,4 +1,5 @@
 add_llvm_library(LLVMRuntimeDyld
   RuntimeDyld.cpp
   RuntimeDyldMachO.cpp
+  RuntimeDyldELF.cpp
   )
diff --git a/lib/ExecutionEngine/RuntimeDyld/RuntimeDyld.cpp b/lib/ExecutionEngine/RuntimeDyld/RuntimeDyld.cpp
index 33dd705027..b017ebb2dc 100644
--- a/lib/ExecutionEngine/RuntimeDyld/RuntimeDyld.cpp
+++ b/lib/ExecutionEngine/RuntimeDyld/RuntimeDyld.cpp
@@ -13,6 +13,7 @@
 
 #define DEBUG_TYPE "dyld"
 #include "RuntimeDyldImpl.h"
+#include "llvm/Support/Path.h"
 using namespace llvm;
 using namespace llvm::object;
 
@@ -64,12 +65,36 @@ RuntimeDyld::~RuntimeDyld() {
 
 bool RuntimeDyld::loadObject(MemoryBuffer *InputBuffer) {
   if (!Dyld) {
-    if (RuntimeDyldMachO::isKnownFormat(InputBuffer))
-      Dyld = new RuntimeDyldMachO(MM);
-    else
-      report_fatal_error("Unknown object format!");
+    sys::LLVMFileType type = sys::IdentifyFileType(
+            InputBuffer->getBufferStart(),
+            static_cast<unsigned>(InputBuffer->getBufferSize()));
+    switch (type) {
+      case sys::ELF_Relocatable_FileType:
+      case sys::ELF_Executable_FileType:
+      case sys::ELF_SharedObject_FileType:
+      case sys::ELF_Core_FileType:
+        Dyld = new RuntimeDyldELF(MM);
+        break;
+      case sys::Mach_O_Object_FileType:
+      case sys::Mach_O_Executable_FileType:
+      case sys::Mach_O_FixedVirtualMemorySharedLib_FileType:
+      case sys::Mach_O_Core_FileType:
+      case sys::Mach_O_PreloadExecutable_FileType:
+      case sys::Mach_O_DynamicallyLinkedSharedLib_FileType:
+      case sys::Mach_O_DynamicLinker_FileType:
+      case sys::Mach_O_Bundle_FileType:
+      case sys::Mach_O_DynamicallyLinkedSharedLibStub_FileType:
+      case sys::Mach_O_DSYMCompanion_FileType:
+        Dyld = new RuntimeDyldMachO(MM);
+        break;
+      case sys::Unknown_FileType:
+      case sys::Bitcode_FileType:
+      case sys::Archive_FileType:
+      case sys::COFF_FileType:
+        report_fatal_error("Incompatible object format!");
+    }
   } else {
-    if(!Dyld->isCompatibleFormat(InputBuffer))
+    if (!Dyld->isCompatibleFormat(InputBuffer))
       report_fatal_error("Incompatible object format!");
   }
 
diff --git a/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp b/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp
new file mode 100644
index 0000000000..9ff95ff8b6
--- /dev/null
+++ b/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp
@@ -0,0 +1,282 @@
+//===-- RuntimeDyldELF.cpp - Run-time dynamic linker for MC-JIT ------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// Implementation of ELF support for the MC-JIT runtime dynamic linker.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "dyld"
+#include "llvm/ADT/OwningPtr.h"
+#include "llvm/ADT/StringRef.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/IntervalMap.h"
+#include "RuntimeDyldImpl.h"
+#include "llvm/Object/ObjectFile.h"
+#include "llvm/Support/ELF.h"
+#include "llvm/ADT/Triple.h"
+using namespace llvm;
+using namespace llvm::object;
+
+namespace llvm {
+
+namespace {
+
+// FIXME: this function should probably not live here...
+//
+// Returns the name and address of an unrelocated symbol in an ELF section
+void getSymbolInfo(symbol_iterator Sym, uint64_t &Addr, StringRef &Name) {
+  //FIXME: error checking here required to catch corrupt ELF objects...
+  error_code Err = Sym->getName(Name);
+
+  uint64_t AddrInSection;
+  Err = Sym->getAddress(AddrInSection);
+
+  SectionRef empty_section;
+  section_iterator Section(empty_section);
+  Err = Sym->getSection(Section);
+
+  StringRef SectionContents;
+  Section->getContents(SectionContents);
+
+  Addr = reinterpret_cast<uint64_t>(SectionContents.data()) + AddrInSection;
+}
+
+}
+
+bool RuntimeDyldELF::loadObject(MemoryBuffer *InputBuffer) {
+  if (!isCompatibleFormat(InputBuffer))
+    return true;
+
+  OwningPtr<ObjectFile> Obj(ObjectFile::createELFObjectFile(InputBuffer));
+
+  Arch = Obj->getArch();
+
+  // Map address in the Object file image to function names
+  IntervalMap<uint64_t, StringRef>::Allocator A;
+  IntervalMap<uint64_t, StringRef> FuncMap(A);
+
+  // This is a bit of a hack.  The ObjectFile we've just loaded reports
+  // section addresses as 0 and doesn't provide access to the section
+  // offset (from which we could calculate the address.  Instead,
+  // we're storing the address when it comes up in the ST_Debug case
+  // below.
+  //
+  StringMap<uint64_t> DebugSymbolMap;
+
+  symbol_iterator SymEnd = Obj->end_symbols();
+  error_code Err;
+  for (symbol_iterator Sym = Obj->begin_symbols();
+       Sym != SymEnd; Sym.increment(Err)) {
+    SymbolRef::Type Type;
+    Sym->getType(Type);
+    if (Type == SymbolRef::ST_Function) {
+      StringRef Name;
+      uint64_t Addr;
+      getSymbolInfo(Sym, Addr, Name);
+
+      uint64_t Size;
+      Err = Sym->getSize(Size);
+
+      uint8_t *Start;
+      uint8_t *End;
+      Start = reinterpret_cast<uint8_t*>(Addr);
+      End   = reinterpret_cast<uint8_t*>(Addr + Size - 1);
+
+      extractFunction(Name, Start, End);
+      FuncMap.insert(Addr, Addr + Size - 1, Name);
+    } else if (Type == SymbolRef::ST_Debug) {
+      // This case helps us find section addresses
+      StringRef Name;
+      uint64_t Addr;
+      getSymbolInfo(Sym, Addr, Name);
+      DebugSymbolMap[Name] = Addr;
+    }
+  }
+
+  // Iterate through the relocations for this object
+  section_iterator SecEnd = Obj->end_sections();
+  for (section_iterator Sec = Obj->begin_sections();
+       Sec != SecEnd; Sec.increment(Err)) {
+    StringRef SecName;
+    uint64_t  SecAddr;
+    Sec->getName(SecName);
+    // Ignore sections that aren't in our map
+    if (DebugSymbolMap.find(SecName) == DebugSymbolMap.end()) {
+      continue;
+    }
+    SecAddr = DebugSymbolMap[SecName];
+    relocation_iterator RelEnd = Sec->end_relocations();
+    for (relocation_iterator Rel = Sec->begin_relocations();
+         Rel != RelEnd; Rel.increment(Err)) {
+      uint64_t RelOffset;
+      uint64_t RelType;
+      int64_t RelAddend;
+      SymbolRef RelSym;
+      StringRef SymName;
+      uint64_t SymAddr;
+      uint64_t SymOffset;
+
+      Rel->getAddress(RelOffset);
+      Rel->getType(RelType);
+      Rel->getAdditionalInfo(RelAddend);
+      Rel->getSymbol(RelSym);
+      RelSym.getName(SymName);
+      RelSym.getAddress(SymAddr);
+      RelSym.getFileOffset(SymOffset);
+
+      // If this relocation is inside a function, we want to store the
+      // function name and a function-relative offset
+      IntervalMap<uint64_t, StringRef>::iterator ContainingFunc
+        = FuncMap.find(SecAddr + RelOffset);
+      if (ContainingFunc.valid()) {
+        // Re-base the relocation to make it relative to the target function
+        RelOffset = (SecAddr + RelOffset) - ContainingFunc.start();
+        Relocations[SymName].push_back(RelocationEntry(ContainingFunc.value(),
+                                                       RelOffset,
+                                                       RelType,
+                                                       RelAddend,
+                                                       true));
+      } else {
+        Relocations[SymName].push_back(RelocationEntry(SecName,
+                                                       RelOffset,
+                                                       RelType,
+                                                       RelAddend,
+                                                       false));
+      }
+    }
+  }
+  return false;
+}
+
+void RuntimeDyldELF::resolveX86_64Relocation(StringRef Name,
+                                             uint8_t *Addr,
+                                             const RelocationEntry &RE) {
+  uint8_t *TargetAddr;
+  if (RE.IsFunctionRelative) {
+    StringMap<sys::MemoryBlock>::iterator ContainingFunc
+      = Functions.find(RE.Target);
+    assert(ContainingFunc != Functions.end()
+           && "Function for relocation not found");
+    TargetAddr = reinterpret_cast<uint8_t*>(ContainingFunc->getValue().base()) +
+                 RE.Offset;
+  } else {
+    // FIXME: Get the address of the target section and add that to RE.Offset
+    assert(0 && ("Non-function relocation not implemented yet!"));
+  }
+
+  switch (RE.Type) {
+  default:
+    assert(0 && ("Relocation type not implemented yet!"));
+  break;
+  case ELF::R_X86_64_64: {
+    uint8_t **Target = reinterpret_cast<uint8_t**>(TargetAddr);
+    *Target = Addr + RE.Addend;
+    break;
+  }
+  case ELF::R_X86_64_32:
+  case ELF::R_X86_64_32S: {
+    uint64_t Value = reinterpret_cast<uint64_t>(Addr) + RE.Addend;
+    // FIXME: Handle the possibility of this assertion failing
+    assert((RE.Type == ELF::R_X86_64_32 && !(Value & 0xFFFFFFFF00000000)) ||
+           (RE.Type == ELF::R_X86_64_32S &&
+            (Value & 0xFFFFFFFF00000000) == 0xFFFFFFFF00000000));
+    uint32_t TruncatedAddr = (Value & 0xFFFFFFFF);
+    uint32_t *Target = reinterpret_cast<uint32_t*>(TargetAddr);
+    *Target = TruncatedAddr;
+    break;
+  }
+  case ELF::R_X86_64_PC32: {
+    uint32_t *Placeholder = reinterpret_cast<uint32_t*>(TargetAddr);
+    uint64_t RealOffset = *Placeholder +
+                           reinterpret_cast<uint64_t>(Addr) +
+                           RE.Addend - reinterpret_cast<uint64_t>(TargetAddr);
+    assert((RealOffset & 0xFFFFFFFF) == RealOffset);
+    uint32_t TruncOffset = (RealOffset & 0xFFFFFFFF);
+    *Placeholder = TruncOffset;
+    break;
+  }
+  }
+}
+
+void RuntimeDyldELF::resolveX86Relocation(StringRef Name,
+                                          uint8_t *Addr,
+                                          const RelocationEntry &RE) {
+  uint8_t *TargetAddr;
+  if (RE.IsFunctionRelative) {
+    StringMap<sys::MemoryBlock>::iterator ContainingFunc
+      = Functions.find(RE.Target);
+    assert(ContainingFunc != Functions.end()
+           && "Function for relocation not found");
+    TargetAddr = reinterpret_cast<uint8_t*>(
+      ContainingFunc->getValue().base()) + RE.Offset;
+  } else {
+    // FIXME: Get the address of the target section and add that to RE.Offset
+    assert(0 && ("Non-function relocation not implemented yet!"));
+  }
+
+  switch (RE.Type) {
+  case ELF::R_386_32: {
+    uint8_t **Target = reinterpret_cast<uint8_t**>(TargetAddr);
+    *Target = Addr + RE.Addend;
+    break;
+  }
+  case ELF::R_386_PC32: {
+    uint32_t *Placeholder = reinterpret_cast<uint32_t*>(TargetAddr);
+    uint32_t RealOffset = *Placeholder + reinterpret_cast<uintptr_t>(Addr) +
+                           RE.Addend - reinterpret_cast<uintptr_t>(TargetAddr);
+    *Placeholder = RealOffset;
+    break;
+    }
+    default:
+      // There are other relocation types, but it appears these are the
+      //  only ones currently used by the LLVM ELF object writer
+      assert(0 && ("Relocation type not implemented yet!"));
+      break;
+  }
+}
+
+void RuntimeDyldELF::resolveArmRelocation(StringRef Name,
+                                          uint8_t *Addr,
+                                          const RelocationEntry &RE) {
+}
+
+void RuntimeDyldELF::resolveRelocation(StringRef Name,
+                                       uint8_t *Addr,
+                                       const RelocationEntry &RE) {
+  switch (Arch) {
+  case Triple::x86_64:
+    resolveX86_64Relocation(Name, Addr, RE);
+    break;
+  case Triple::x86:
+    resolveX86Relocation(Name, Addr, RE);
+    break;
+  case Triple::arm:
+    resolveArmRelocation(Name, Addr, RE);
+    break;
+  default:
+    assert(0 && "Unsupported CPU type!");
+    break;
+  }
+}
+
+void RuntimeDyldELF::reassignSymbolAddress(StringRef Name, uint8_t *Addr) {
+  SymbolTable[Name] = Addr;
+
+  RelocationList &Relocs = Relocations[Name];
+  for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
+    RelocationEntry &RE = Relocs[i];
+    resolveRelocation(Name, Addr, RE);
+  }
+}
+
+bool RuntimeDyldELF::isCompatibleFormat(const MemoryBuffer *InputBuffer) const {
+  StringRef Magic = InputBuffer->getBuffer().slice(0, ELF::EI_NIDENT);
+  return (memcmp(Magic.data(), ELF::ElfMagic, strlen(ELF::ElfMagic))) == 0;
+}
+} // namespace llvm
diff --git a/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldImpl.h b/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldImpl.h
index 7190a3c36f..cff7cbdf28 100644
--- a/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldImpl.h
+++ b/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldImpl.h
@@ -94,6 +94,66 @@ public:
   virtual bool isCompatibleFormat(const MemoryBuffer *InputBuffer) const = 0;
 };
 
+class RuntimeDyldELF : public RuntimeDyldImpl {
+    // For each symbol, keep a list of relocations based on it. Anytime
+    // its address is reassigned (the JIT re-compiled the function, e.g.),
+    // the relocations get re-resolved.
+    struct RelocationEntry {
+      // Function or section this relocation is contained in.
+      std::string Target;
+      // Offset into the target function or section for the relocation.
+      uint32_t    Offset;
+      // Relocation type
+      uint32_t    Type;
+      // Addend encoded in the instruction itself, if any.
+      int32_t     Addend;
+      // Has the relocation been recalcuated as an offset within a function?
+      bool        IsFunctionRelative;
+      // Has this relocation been resolved previously?
+      bool        isResolved;
+
+      RelocationEntry(StringRef t,
+                      uint32_t offset,
+                      uint32_t type,
+                      int32_t addend,
+                      bool isFunctionRelative)
+        : Target(t)
+        , Offset(offset)
+        , Type(type)
+        , Addend(addend)
+        , IsFunctionRelative(isFunctionRelative)
+        , isResolved(false) { }
+    };
+    typedef SmallVector<RelocationEntry, 4> RelocationList;
+    StringMap<RelocationList> Relocations;
+    unsigned Arch;
+
+    void resolveX86_64Relocation(StringRef Name,
+                                 uint8_t *Addr,
+                                 const RelocationEntry &RE);
+
+    void resolveX86Relocation(StringRef Name,
+                              uint8_t *Addr,
+                              const RelocationEntry &RE);
+
+    void resolveArmRelocation(StringRef Name,
+                              uint8_t *Addr,
+                              const RelocationEntry &RE);
+
+    void resolveRelocation(StringRef Name,
+                           uint8_t *Addr,
+                           const RelocationEntry &RE);
+
+public:
+  RuntimeDyldELF(RTDyldMemoryManager *mm) : RuntimeDyldImpl(mm) {}
+
+  bool loadObject(MemoryBuffer *InputBuffer);
+
+  void reassignSymbolAddress(StringRef Name, uint8_t *Addr);
+
+  bool isCompatibleFormat(const MemoryBuffer *InputBuffer) const;
+};
+
 
 class RuntimeDyldMachO : public RuntimeDyldImpl {