[Orc] New JIT APIs.

This patch adds a new set of JIT APIs to LLVM. The aim of these new APIs is to
cleanly support a wider range of JIT use cases in LLVM, and encourage the
development and contribution of re-usable infrastructure for LLVM JIT use-cases.

These APIs are intended to live alongside the MCJIT APIs, and should not affect
existing clients.

Included in this patch:

1) New headers in include/llvm/ExecutionEngine/Orc that provide a set of
   components for building JIT infrastructure.
   Implementation code for these headers lives in lib/ExecutionEngine/Orc.

2) A prototype re-implementation of MCJIT (OrcMCJITReplacement) built out of the
   new components.

3) Minor changes to RTDyldMemoryManager needed to support the new components.
   These changes should not impact existing clients.

4) A new flag for lli, -use-orcmcjit, which will cause lli to use the
   OrcMCJITReplacement class as its underlying execution engine, rather than
   MCJIT itself.

Tests to follow shortly.

Special thanks to Michael Ilseman, Pete Cooper, David Blaikie, Eric Christopher,
Justin Bogner, and Jim Grosbach for extensive feedback and discussion.

llvm-svn: 226940

8 files changed, 864 insertions, 0 deletions

diff --git a/llvm/lib/ExecutionEngine/Orc/CMakeLists.txt b/llvm/lib/ExecutionEngine/Orc/CMakeLists.txt
new file mode 100644
index 00000000000..3fadb931200
--- /dev/null
+++ b/llvm/lib/ExecutionEngine/Orc/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_llvm_library(LLVMOrcJIT
+  CloneSubModule.cpp
+  IndirectionUtils.cpp
+  OrcMCJITReplacement.cpp
+  OrcTargetSupport.cpp
+  )
diff --git a/llvm/lib/ExecutionEngine/Orc/CloneSubModule.cpp b/llvm/lib/ExecutionEngine/Orc/CloneSubModule.cpp
new file mode 100644
index 00000000000..54acb78153e
--- /dev/null
+++ b/llvm/lib/ExecutionEngine/Orc/CloneSubModule.cpp
@@ -0,0 +1,112 @@
+#include "llvm/ExecutionEngine/Orc/CloneSubModule.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/Module.h"
+#include "llvm/Transforms/Utils/Cloning.h"
+
+using namespace llvm;
+
+void llvm::copyGVInitializer(GlobalVariable &New, const GlobalVariable &Orig,
+                             ValueToValueMapTy &VMap) {
+  if (Orig.hasInitializer())
+    New.setInitializer(MapValue(Orig.getInitializer(), VMap));
+}
+
+void llvm::copyFunctionBody(Function &New, const Function &Orig,
+                            ValueToValueMapTy &VMap) {
+  if (!Orig.isDeclaration()) {
+    Function::arg_iterator DestI = New.arg_begin();
+    for (Function::const_arg_iterator J = Orig.arg_begin(); J != Orig.arg_end();
+         ++J) {
+      DestI->setName(J->getName());
+      VMap[J] = DestI++;
+    }
+
+    SmallVector<ReturnInst *, 8> Returns; // Ignore returns cloned.
+    CloneFunctionInto(&New, &Orig, VMap, /*ModuleLevelChanges=*/true, Returns);
+  }
+}
+
+std::unique_ptr<Module>
+llvm::CloneSubModule(const Module &M,
+                     HandleGlobalVariableFtor HandleGlobalVariable,
+                     HandleFunctionFtor HandleFunction, bool KeepInlineAsm) {
+
+  ValueToValueMapTy VMap;
+
+  // First off, we need to create the new module.
+  std::unique_ptr<Module> New =
+      llvm::make_unique<Module>(M.getModuleIdentifier(), M.getContext());
+
+  New->setDataLayout(M.getDataLayout());
+  New->setTargetTriple(M.getTargetTriple());
+  if (KeepInlineAsm)
+    New->setModuleInlineAsm(M.getModuleInlineAsm());
+
+  // Copy global variables (but not initializers, yet).
+  for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
+       I != E; ++I) {
+    GlobalVariable *GV = new GlobalVariable(
+        *New, I->getType()->getElementType(), I->isConstant(), I->getLinkage(),
+        (Constant *)nullptr, I->getName(), (GlobalVariable *)nullptr,
+        I->getThreadLocalMode(), I->getType()->getAddressSpace());
+    GV->copyAttributesFrom(I);
+    VMap[I] = GV;
+  }
+
+  // Loop over the functions in the module, making external functions as before
+  for (Module::const_iterator I = M.begin(), E = M.end(); I != E; ++I) {
+    Function *NF =
+        Function::Create(cast<FunctionType>(I->getType()->getElementType()),
+                         I->getLinkage(), I->getName(), &*New);
+    NF->copyAttributesFrom(I);
+    VMap[I] = NF;
+  }
+
+  // Loop over the aliases in the module
+  for (Module::const_alias_iterator I = M.alias_begin(), E = M.alias_end();
+       I != E; ++I) {
+    auto *PTy = cast<PointerType>(I->getType());
+    auto *GA =
+        GlobalAlias::create(PTy->getElementType(), PTy->getAddressSpace(),
+                            I->getLinkage(), I->getName(), &*New);
+    GA->copyAttributesFrom(I);
+    VMap[I] = GA;
+  }
+
+  // Now that all of the things that global variable initializer can refer to
+  // have been created, loop through and copy the global variable referrers
+  // over...  We also set the attributes on the global now.
+  for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
+       I != E; ++I) {
+    GlobalVariable &GV = *cast<GlobalVariable>(VMap[I]);
+    HandleGlobalVariable(GV, *I, VMap);
+  }
+
+  // Similarly, copy over function bodies now...
+  //
+  for (Module::const_iterator I = M.begin(), E = M.end(); I != E; ++I) {
+    Function &F = *cast<Function>(VMap[I]);
+    HandleFunction(F, *I, VMap);
+  }
+
+  // And aliases
+  for (Module::const_alias_iterator I = M.alias_begin(), E = M.alias_end();
+       I != E; ++I) {
+    GlobalAlias *GA = cast<GlobalAlias>(VMap[I]);
+    if (const Constant *C = I->getAliasee())
+      GA->setAliasee(MapValue(C, VMap));
+  }
+
+  // And named metadata....
+  for (Module::const_named_metadata_iterator I = M.named_metadata_begin(),
+                                             E = M.named_metadata_end();
+       I != E; ++I) {
+    const NamedMDNode &NMD = *I;
+    NamedMDNode *NewNMD = New->getOrInsertNamedMetadata(NMD.getName());
+    for (unsigned i = 0, e = NMD.getNumOperands(); i != e; ++i)
+      NewNMD->addOperand(MapMetadata(NMD.getOperand(i), VMap));
+  }
+
+  return New;
+}
diff --git a/llvm/lib/ExecutionEngine/Orc/IndirectionUtils.cpp b/llvm/lib/ExecutionEngine/Orc/IndirectionUtils.cpp
new file mode 100644
index 00000000000..e75092b2485
--- /dev/null
+++ b/llvm/lib/ExecutionEngine/Orc/IndirectionUtils.cpp
@@ -0,0 +1,157 @@
+#include "llvm/ADT/Triple.h"
+#include "llvm/ExecutionEngine/Orc/IndirectionUtils.h"
+#include "llvm/ExecutionEngine/Orc/CloneSubModule.h"
+#include "llvm/IR/CallSite.h"
+#include "llvm/IR/IRBuilder.h"
+#include <set>
+
+using namespace llvm;
+
+namespace llvm {
+
+JITIndirections makeCallsSingleIndirect(
+    Module &M, const std::function<bool(const Function &)> &ShouldIndirect,
+    const char *JITImplSuffix, const char *JITAddrSuffix) {
+  std::vector<Function *> Worklist;
+  std::vector<std::string> FuncNames;
+
+  for (auto &F : M)
+    if (ShouldIndirect(F) && (F.user_begin() != F.user_end())) {
+      Worklist.push_back(&F);
+      FuncNames.push_back(F.getName());
+    }
+
+  for (auto *F : Worklist) {
+    GlobalVariable *FImplAddr = new GlobalVariable(
+        M, F->getType(), false, GlobalValue::ExternalLinkage,
+        Constant::getNullValue(F->getType()), F->getName() + JITAddrSuffix,
+        nullptr, GlobalValue::NotThreadLocal, 0, true);
+    FImplAddr->setVisibility(GlobalValue::HiddenVisibility);
+
+    for (auto *U : F->users()) {
+      assert(isa<Instruction>(U) && "Cannot indirect non-instruction use");
+      IRBuilder<> Builder(cast<Instruction>(U));
+      U->replaceUsesOfWith(F, Builder.CreateLoad(FImplAddr));
+    }
+  }
+
+  return JITIndirections(
+      FuncNames, [=](StringRef S) -> std::string { return std::string(S); },
+      [=](StringRef S)
+          -> std::string { return std::string(S) + JITAddrSuffix; });
+}
+
+JITIndirections makeCallsDoubleIndirect(
+    Module &M, const std::function<bool(const Function &)> &ShouldIndirect,
+    const char *JITImplSuffix, const char *JITAddrSuffix) {
+
+  std::vector<Function *> Worklist;
+  std::vector<std::string> FuncNames;
+
+  for (auto &F : M)
+    if (!F.isDeclaration() && !F.hasAvailableExternallyLinkage() &&
+        ShouldIndirect(F))
+      Worklist.push_back(&F);
+
+  for (auto *F : Worklist) {
+    std::string OrigName = F->getName();
+    F->setName(OrigName + JITImplSuffix);
+    FuncNames.push_back(OrigName);
+
+    GlobalVariable *FImplAddr = new GlobalVariable(
+        M, F->getType(), false, GlobalValue::ExternalLinkage,
+        Constant::getNullValue(F->getType()), OrigName + JITAddrSuffix, nullptr,
+        GlobalValue::NotThreadLocal, 0, true);
+    FImplAddr->setVisibility(GlobalValue::HiddenVisibility);
+
+    Function *FRedirect =
+        Function::Create(F->getFunctionType(), F->getLinkage(), OrigName, &M);
+
+    F->replaceAllUsesWith(FRedirect);
+
+    BasicBlock *EntryBlock =
+        BasicBlock::Create(M.getContext(), "entry", FRedirect);
+
+    IRBuilder<> Builder(EntryBlock);
+    LoadInst *FImplLoadedAddr = Builder.CreateLoad(FImplAddr);
+
+    std::vector<Value *> CallArgs;
+    for (Value &Arg : FRedirect->args())
+      CallArgs.push_back(&Arg);
+    CallInst *Call = Builder.CreateCall(FImplLoadedAddr, CallArgs);
+    Call->setTailCall();
+    Builder.CreateRet(Call);
+  }
+
+  return JITIndirections(
+      FuncNames, [=](StringRef S)
+                     -> std::string { return std::string(S) + JITImplSuffix; },
+      [=](StringRef S)
+          -> std::string { return std::string(S) + JITAddrSuffix; });
+}
+
+std::vector<std::unique_ptr<Module>>
+explode(const Module &OrigMod,
+        const std::function<bool(const Function &)> &ShouldExtract) {
+
+  std::vector<std::unique_ptr<Module>> NewModules;
+
+  // Split all the globals, non-indirected functions, etc. into a single module.
+  auto ExtractGlobalVars = [&](GlobalVariable &New, const GlobalVariable &Orig,
+                               ValueToValueMapTy &VMap) {
+    copyGVInitializer(New, Orig, VMap);
+    if (New.getLinkage() == GlobalValue::PrivateLinkage) {
+      New.setLinkage(GlobalValue::ExternalLinkage);
+      New.setVisibility(GlobalValue::HiddenVisibility);
+    }
+  };
+
+  auto ExtractNonImplFunctions =
+      [&](Function &New, const Function &Orig, ValueToValueMapTy &VMap) {
+        if (!ShouldExtract(New))
+          copyFunctionBody(New, Orig, VMap);
+      };
+
+  NewModules.push_back(CloneSubModule(OrigMod, ExtractGlobalVars,
+                                      ExtractNonImplFunctions, true));
+
+  // Preserve initializers for Common linkage vars, and make private linkage
+  // globals external: they are now provided by the globals module extracted
+  // above.
+  auto DropGlobalVars = [&](GlobalVariable &New, const GlobalVariable &Orig,
+                            ValueToValueMapTy &VMap) {
+    if (New.getLinkage() == GlobalValue::CommonLinkage)
+      copyGVInitializer(New, Orig, VMap);
+    else if (New.getLinkage() == GlobalValue::PrivateLinkage)
+      New.setLinkage(GlobalValue::ExternalLinkage);
+  };
+
+  // Split each 'impl' function out in to its own module.
+  for (const auto &Func : OrigMod) {
+    if (Func.isDeclaration() || !ShouldExtract(Func))
+      continue;
+
+    auto ExtractNamedFunction =
+        [&](Function &New, const Function &Orig, ValueToValueMapTy &VMap) {
+          if (New.getName() == Func.getName())
+            copyFunctionBody(New, Orig, VMap);
+        };
+
+    NewModules.push_back(
+        CloneSubModule(OrigMod, DropGlobalVars, ExtractNamedFunction, false));
+  }
+
+  return NewModules;
+}
+
+std::vector<std::unique_ptr<Module>>
+explode(const Module &OrigMod, const JITIndirections &Indirections) {
+  std::set<std::string> ImplNames;
+
+  for (const auto &FuncName : Indirections.IndirectedNames)
+    ImplNames.insert(Indirections.GetImplName(FuncName));
+
+  return explode(
+      OrigMod, [&](const Function &F) { return ImplNames.count(F.getName()); });
+}
+}
diff --git a/llvm/lib/ExecutionEngine/Orc/LLVMBuild.txt b/llvm/lib/ExecutionEngine/Orc/LLVMBuild.txt
new file mode 100644
index 00000000000..620449405b5
--- /dev/null
+++ b/llvm/lib/ExecutionEngine/Orc/LLVMBuild.txt
@@ -0,0 +1,22 @@
+;===- ./lib/ExecutionEngine/MCJIT/LLVMBuild.txt ----------------*- Conf -*--===;
+;
+;                     The LLVM Compiler Infrastructure
+;
+; This file is distributed under the University of Illinois Open Source
+; License. See LICENSE.TXT for details.
+;
+;===------------------------------------------------------------------------===;
+;
+; This is an LLVMBuild description file for the components in this subdirectory.
+;
+; For more information on the LLVMBuild system, please see:
+;
+;   http://llvm.org/docs/LLVMBuild.html
+;
+;===------------------------------------------------------------------------===;
+
+[component_0]
+type = Library
+name = OrcJIT
+parent = ExecutionEngine
+required_libraries = Core ExecutionEngine Object RuntimeDyld Support Target
diff --git a/llvm/lib/ExecutionEngine/Orc/Makefile b/llvm/lib/ExecutionEngine/Orc/Makefile
new file mode 100644
index 00000000000..ac302348ee7
--- /dev/null
+++ b/llvm/lib/ExecutionEngine/Orc/Makefile
@@ -0,0 +1,13 @@
+##===- lib/ExecutionEngine/OrcJIT/Makefile -----------------*- Makefile -*-===##
+#
+#                     The LLVM Compiler Infrastructure
+#
+# This file is distributed under the University of Illinois Open Source
+# License. See LICENSE.TXT for details.
+#
+##===----------------------------------------------------------------------===##
+
+LEVEL = ../../..
+LIBRARYNAME = LLVMOrcJIT
+
+include $(LEVEL)/Makefile.common
diff --git a/llvm/lib/ExecutionEngine/Orc/OrcMCJITReplacement.cpp b/llvm/lib/ExecutionEngine/Orc/OrcMCJITReplacement.cpp
new file mode 100644
index 00000000000..f658ab2b298
--- /dev/null
+++ b/llvm/lib/ExecutionEngine/Orc/OrcMCJITReplacement.cpp
@@ -0,0 +1,124 @@
+//===-------- OrcMCJITReplacement.cpp - Orc-based MCJIT replacement -------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#include "OrcMCJITReplacement.h"
+#include "llvm/ExecutionEngine/GenericValue.h"
+
+namespace {
+
+static struct RegisterJIT {
+  RegisterJIT() { llvm::OrcMCJITReplacement::Register(); }
+} JITRegistrator;
+
+extern "C" void LLVMLinkInOrcMCJITReplacement() {}
+}
+
+namespace llvm {
+
+GenericValue
+OrcMCJITReplacement::runFunction(Function *F,
+                                 const std::vector<GenericValue> &ArgValues) {
+  assert(F && "Function *F was null at entry to run()");
+
+  void *FPtr = getPointerToFunction(F);
+  assert(FPtr && "Pointer to fn's code was null after getPointerToFunction");
+  FunctionType *FTy = F->getFunctionType();
+  Type *RetTy = FTy->getReturnType();
+
+  assert((FTy->getNumParams() == ArgValues.size() ||
+          (FTy->isVarArg() && FTy->getNumParams() <= ArgValues.size())) &&
+         "Wrong number of arguments passed into function!");
+  assert(FTy->getNumParams() == ArgValues.size() &&
+         "This doesn't support passing arguments through varargs (yet)!");
+
+  // Handle some common cases first.  These cases correspond to common `main'
+  // prototypes.
+  if (RetTy->isIntegerTy(32) || RetTy->isVoidTy()) {
+    switch (ArgValues.size()) {
+    case 3:
+      if (FTy->getParamType(0)->isIntegerTy(32) &&
+          FTy->getParamType(1)->isPointerTy() &&
+          FTy->getParamType(2)->isPointerTy()) {
+        int (*PF)(int, char **, const char **) =
+            (int (*)(int, char **, const char **))(intptr_t)FPtr;
+
+        // Call the function.
+        GenericValue rv;
+        rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue(),
+                                 (char **)GVTOP(ArgValues[1]),
+                                 (const char **)GVTOP(ArgValues[2])));
+        return rv;
+      }
+      break;
+    case 2:
+      if (FTy->getParamType(0)->isIntegerTy(32) &&
+          FTy->getParamType(1)->isPointerTy()) {
+        int (*PF)(int, char **) = (int (*)(int, char **))(intptr_t)FPtr;
+
+        // Call the function.
+        GenericValue rv;
+        rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue(),
+                                 (char **)GVTOP(ArgValues[1])));
+        return rv;
+      }
+      break;
+    case 1:
+      if (FTy->getNumParams() == 1 && FTy->getParamType(0)->isIntegerTy(32)) {
+        GenericValue rv;
+        int (*PF)(int) = (int (*)(int))(intptr_t)FPtr;
+        rv.IntVal = APInt(32, PF(ArgValues[0].IntVal.getZExtValue()));
+        return rv;
+      }
+      break;
+    }
+  }
+
+  // Handle cases where no arguments are passed first.
+  if (ArgValues.empty()) {
+    GenericValue rv;
+    switch (RetTy->getTypeID()) {
+    default:
+      llvm_unreachable("Unknown return type for function call!");
+    case Type::IntegerTyID: {
+      unsigned BitWidth = cast<IntegerType>(RetTy)->getBitWidth();
+      if (BitWidth == 1)
+        rv.IntVal = APInt(BitWidth, ((bool (*)())(intptr_t)FPtr)());
+      else if (BitWidth <= 8)
+        rv.IntVal = APInt(BitWidth, ((char (*)())(intptr_t)FPtr)());
+      else if (BitWidth <= 16)
+        rv.IntVal = APInt(BitWidth, ((short (*)())(intptr_t)FPtr)());
+      else if (BitWidth <= 32)
+        rv.IntVal = APInt(BitWidth, ((int (*)())(intptr_t)FPtr)());
+      else if (BitWidth <= 64)
+        rv.IntVal = APInt(BitWidth, ((int64_t (*)())(intptr_t)FPtr)());
+      else
+        llvm_unreachable("Integer types > 64 bits not supported");
+      return rv;
+    }
+    case Type::VoidTyID:
+      rv.IntVal = APInt(32, ((int (*)())(intptr_t)FPtr)());
+      return rv;
+    case Type::FloatTyID:
+      rv.FloatVal = ((float (*)())(intptr_t)FPtr)();
+      return rv;
+    case Type::DoubleTyID:
+      rv.DoubleVal = ((double (*)())(intptr_t)FPtr)();
+      return rv;
+    case Type::X86_FP80TyID:
+    case Type::FP128TyID:
+    case Type::PPC_FP128TyID:
+      llvm_unreachable("long double not supported yet");
+    case Type::PointerTyID:
+      return PTOGV(((void *(*)())(intptr_t)FPtr)());
+    }
+  }
+
+  llvm_unreachable("Full-featured argument passing not supported yet!");
+}
+}
diff --git a/llvm/lib/ExecutionEngine/Orc/OrcMCJITReplacement.h b/llvm/lib/ExecutionEngine/Orc/OrcMCJITReplacement.h
new file mode 100644
index 00000000000..9fdf0efdd90
--- /dev/null
+++ b/llvm/lib/ExecutionEngine/Orc/OrcMCJITReplacement.h
@@ -0,0 +1,328 @@
+//===---- OrcMCJITReplacement.h - Orc based MCJIT replacement ---*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// Orc based MCJIT replacement.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_EXECUTIONENGINE_ORC_ORCMCJITREPLACEMENT_H
+#define LLVM_LIB_EXECUTIONENGINE_ORC_ORCMCJITREPLACEMENT_H
+
+#include "llvm/ExecutionEngine/ExecutionEngine.h"
+#include "llvm/ExecutionEngine/Orc/CompileUtils.h"
+#include "llvm/ExecutionEngine/Orc/IRCompileLayer.h"
+#include "llvm/ExecutionEngine/Orc/LazyEmittingLayer.h"
+#include "llvm/ExecutionEngine/Orc/ObjectLinkingLayer.h"
+#include "llvm/Object/Archive.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
+
+namespace llvm {
+
+class OrcMCJITReplacement : public ExecutionEngine {
+
+  class ForwardingRTDyldMM : public RTDyldMemoryManager {
+  public:
+    ForwardingRTDyldMM(OrcMCJITReplacement &M) : M(M) {}
+
+    uint8_t *allocateCodeSection(uintptr_t Size, unsigned Alignment,
+                                 unsigned SectionID,
+                                 StringRef SectionName) override {
+      uint8_t *Addr =
+          M.MM->allocateCodeSection(Size, Alignment, SectionID, SectionName);
+      M.SectionsAllocatedSinceLastLoad.insert(Addr);
+      return Addr;
+    }
+
+    uint8_t *allocateDataSection(uintptr_t Size, unsigned Alignment,
+                                 unsigned SectionID, StringRef SectionName,
+                                 bool IsReadOnly) override {
+      uint8_t *Addr = M.MM->allocateDataSection(Size, Alignment, SectionID,
+                                                SectionName, IsReadOnly);
+      M.SectionsAllocatedSinceLastLoad.insert(Addr);
+      return Addr;
+    }
+
+    void reserveAllocationSpace(uintptr_t CodeSize, uintptr_t DataSizeRO,
+                                uintptr_t DataSizeRW) override {
+      return M.MM->reserveAllocationSpace(CodeSize, DataSizeRO, DataSizeRW);
+    }
+
+    bool needsToReserveAllocationSpace() override {
+      return M.MM->needsToReserveAllocationSpace();
+    }
+
+    void registerEHFrames(uint8_t *Addr, uint64_t LoadAddr,
+                          size_t Size) override {
+      return M.MM->registerEHFrames(Addr, LoadAddr, Size);
+    }
+
+    void deregisterEHFrames(uint8_t *Addr, uint64_t LoadAddr,
+                            size_t Size) override {
+      return M.MM->deregisterEHFrames(Addr, LoadAddr, Size);
+    }
+
+    uint64_t getSymbolAddress(const std::string &Name) override {
+      return M.getSymbolAddressWithoutMangling(Name);
+    }
+
+    void *getPointerToNamedFunction(const std::string &Name,
+                                    bool AbortOnFailure = true) override {
+      return M.MM->getPointerToNamedFunction(Name, AbortOnFailure);
+    }
+
+    void notifyObjectLoaded(ExecutionEngine *EE,
+                            const object::ObjectFile &O) override {
+      return M.MM->notifyObjectLoaded(EE, O);
+    }
+
+    bool finalizeMemory(std::string *ErrMsg = nullptr) override {
+      // Each set of objects loaded will be finalized exactly once, but since
+      // symbol lookup during relocation may recursively trigger the
+      // loading/relocation of other modules, and since we're forwarding all
+      // finalizeMemory calls to a single underlying memory manager, we need to
+      // defer forwarding the call on until all necessary objects have been
+      // loaded. Otherwise, during the relocation of a leaf object, we will end
+      // up finalizing memory, causing a crash further up the stack when we
+      // attempt to apply relocations to finalized memory.
+      // To avoid finalizing too early, look at how many objects have been
+      // loaded but not yet finalized. This is a bit of a hack that relies on
+      // the fact that we're lazily emitting object files: The only way you can
+      // get more than one set of objects loaded but not yet finalized is if
+      // they were loaded during relocation of another set.
+      if (M.UnfinalizedSections.size() == 1)
+        return M.MM->finalizeMemory(ErrMsg);
+      return false;
+    }
+
+  private:
+    OrcMCJITReplacement &M;
+  };
+
+private:
+  static ExecutionEngine *
+  createOrcMCJITReplacement(std::string *ErrorMsg,
+                            std::unique_ptr<RTDyldMemoryManager> OrcJMM,
+                            std::unique_ptr<llvm::TargetMachine> TM) {
+    return new llvm::OrcMCJITReplacement(std::move(OrcJMM), std::move(TM));
+  }
+
+public:
+  static void Register() {
+    OrcMCJITReplacementCtor = createOrcMCJITReplacement;
+  }
+
+  OrcMCJITReplacement(std::unique_ptr<RTDyldMemoryManager> MM,
+                      std::unique_ptr<TargetMachine> TM)
+      : TM(std::move(TM)), MM(std::move(MM)),
+        Mang(this->TM->getSubtargetImpl()->getDataLayout()),
+        NotifyObjectLoaded(*this), NotifyFinalized(*this),
+        ObjectLayer(NotifyObjectLoaded, NotifyFinalized),
+        CompileLayer(ObjectLayer, SimpleCompiler(*this->TM)),
+        LazyEmitLayer(CompileLayer) {
+    setDataLayout(this->TM->getSubtargetImpl()->getDataLayout());
+  }
+
+  void addModule(std::unique_ptr<Module> M) {
+
+    // If this module doesn't have a DataLayout attached then attach the
+    // default.
+    if (!M->getDataLayout())
+      M->setDataLayout(getDataLayout());
+
+    OwnedModules.push_back(std::move(M));
+    std::vector<Module *> Ms;
+    Ms.push_back(&*OwnedModules.back());
+    LazyEmitLayer.addModuleSet(std::move(Ms),
+                               llvm::make_unique<ForwardingRTDyldMM>(*this));
+  }
+
+  void addObjectFile(std::unique_ptr<object::ObjectFile> O) override {
+    std::vector<std::unique_ptr<object::ObjectFile>> Objs;
+    Objs.push_back(std::move(O));
+    ObjectLayer.addObjectSet(std::move(Objs),
+                             llvm::make_unique<ForwardingRTDyldMM>(*this));
+  }
+
+  void addObjectFile(object::OwningBinary<object::ObjectFile> O) override {
+    std::unique_ptr<object::ObjectFile> Obj;
+    std::unique_ptr<MemoryBuffer> Buf;
+    std::tie(Obj, Buf) = O.takeBinary();
+    std::vector<std::unique_ptr<object::ObjectFile>> Objs;
+    Objs.push_back(std::move(Obj));
+    ObjectLayer.addObjectSet(std::move(Objs),
+                             llvm::make_unique<ForwardingRTDyldMM>(*this));
+  }
+
+  void addArchive(object::OwningBinary<object::Archive> A) override {
+    Archives.push_back(std::move(A));
+  }
+
+  uint64_t getSymbolAddress(StringRef Name) {
+    return getSymbolAddressWithoutMangling(Mangle(Name));
+  }
+
+  void finalizeObject() override {
+    // This is deprecated - Aim to remove in ExecutionEngine.
+    // REMOVE IF POSSIBLE - Doesn't make sense for New JIT.
+  }
+
+  void mapSectionAddress(const void *LocalAddress,
+                         uint64_t TargetAddress) override {
+    for (auto &P : UnfinalizedSections)
+      if (P.second.count(LocalAddress))
+        ObjectLayer.mapSectionAddress(P.first, LocalAddress, TargetAddress);
+  }
+
+  uint64_t getGlobalValueAddress(const std::string &Name) override {
+    return getSymbolAddress(Name);
+  }
+
+  uint64_t getFunctionAddress(const std::string &Name) override {
+    return getSymbolAddress(Name);
+  }
+
+  void *getPointerToFunction(Function *F) override {
+    uint64_t FAddr = getSymbolAddress(F->getName());
+    return reinterpret_cast<void *>(static_cast<uintptr_t>(FAddr));
+  }
+
+  void *getPointerToNamedFunction(StringRef Name,
+                                  bool AbortOnFailure = true) override {
+    uint64_t Addr = getSymbolAddress(Name);
+    if (!Addr && AbortOnFailure)
+      llvm_unreachable("Missing symbol!");
+    return reinterpret_cast<void *>(static_cast<uintptr_t>(Addr));
+  }
+
+  GenericValue runFunction(Function *F,
+                           const std::vector<GenericValue> &ArgValues) override;
+
+  void setObjectCache(ObjectCache *NewCache) override {
+    CompileLayer.setObjectCache(NewCache);
+  }
+
+private:
+  uint64_t getSymbolAddressWithoutMangling(StringRef Name) {
+    if (uint64_t Addr = LazyEmitLayer.getSymbolAddress(Name, false))
+      return Addr;
+    if (uint64_t Addr = MM->getSymbolAddress(Name))
+      return Addr;
+    if (uint64_t Addr = scanArchives(Name))
+      return Addr;
+
+    return 0;
+  }
+
+  uint64_t scanArchives(StringRef Name) {
+    for (object::OwningBinary<object::Archive> &OB : Archives) {
+      object::Archive *A = OB.getBinary();
+      // Look for our symbols in each Archive
+      object::Archive::child_iterator ChildIt = A->findSym(Name);
+      if (ChildIt != A->child_end()) {
+        // FIXME: Support nested archives?
+        ErrorOr<std::unique_ptr<object::Binary>> ChildBinOrErr =
+            ChildIt->getAsBinary();
+        if (ChildBinOrErr.getError())
+          continue;
+        std::unique_ptr<object::Binary> &ChildBin = ChildBinOrErr.get();
+        if (ChildBin->isObject()) {
+          std::vector<std::unique_ptr<object::ObjectFile>> ObjSet;
+          ObjSet.push_back(std::unique_ptr<object::ObjectFile>(
+              static_cast<object::ObjectFile *>(ChildBin.release())));
+          ObjectLayer.addObjectSet(
+              std::move(ObjSet), llvm::make_unique<ForwardingRTDyldMM>(*this));
+          if (uint64_t Addr = ObjectLayer.getSymbolAddress(Name, true))
+            return Addr;
+        }
+      }
+    }
+    return 0;
+  }
+
+  class NotifyObjectLoadedT {
+  public:
+    typedef std::vector<std::unique_ptr<object::ObjectFile>> ObjListT;
+    typedef std::vector<std::unique_ptr<RuntimeDyld::LoadedObjectInfo>>
+        LoadedObjInfoListT;
+
+    NotifyObjectLoadedT(OrcMCJITReplacement &M) : M(M) {}
+
+    void operator()(ObjectLinkingLayerBase::ObjSetHandleT H,
+                    const ObjListT &Objects,
+                    const LoadedObjInfoListT &Infos) const {
+      M.UnfinalizedSections[H] = std::move(M.SectionsAllocatedSinceLastLoad);
+      M.SectionsAllocatedSinceLastLoad = {};
+      assert(Objects.size() == Infos.size() &&
+             "Incorrect number of Infos for Objects.");
+      for (unsigned I = 0; I < Objects.size(); ++I)
+        M.MM->notifyObjectLoaded(&M, *Objects[I]);
+    };
+
+  private:
+    OrcMCJITReplacement &M;
+  };
+
+  class NotifyFinalizedT {
+  public:
+    NotifyFinalizedT(OrcMCJITReplacement &M) : M(M) {}
+    void operator()(ObjectLinkingLayerBase::ObjSetHandleT H) {
+      M.UnfinalizedSections.erase(H);
+    }
+
+  private:
+    OrcMCJITReplacement &M;
+  };
+
+  std::string Mangle(StringRef Name) {
+    std::string MangledName;
+    {
+      raw_string_ostream MangledNameStream(MangledName);
+      Mang.getNameWithPrefix(MangledNameStream, Name);
+    }
+    return MangledName;
+  }
+
+  typedef ObjectLinkingLayer<NotifyObjectLoadedT> ObjectLayerT;
+  typedef IRCompileLayer<ObjectLayerT> CompileLayerT;
+  typedef LazyEmittingLayer<CompileLayerT> LazyEmitLayerT;
+
+  std::unique_ptr<TargetMachine> TM;
+  std::unique_ptr<RTDyldMemoryManager> MM;
+  Mangler Mang;
+
+  NotifyObjectLoadedT NotifyObjectLoaded;
+  NotifyFinalizedT NotifyFinalized;
+
+  ObjectLayerT ObjectLayer;
+  CompileLayerT CompileLayer;
+  LazyEmitLayerT LazyEmitLayer;
+
+  // MCJIT keeps modules alive - we need to do the same for backwards
+  // compatibility.
+  std::vector<std::unique_ptr<Module>> OwnedModules;
+
+  // We need to store ObjLayerT::ObjSetHandles for each of the object sets
+  // that have been emitted but not yet finalized so that we can forward the
+  // mapSectionAddress calls appropriately.
+  typedef std::set<const void *> SectionAddrSet;
+  struct ObjSetHandleCompare {
+    bool operator()(ObjectLayerT::ObjSetHandleT H1,
+                    ObjectLayerT::ObjSetHandleT H2) const {
+      return &*H1 < &*H2;
+    }
+  };
+  SectionAddrSet SectionsAllocatedSinceLastLoad;
+  std::map<ObjectLayerT::ObjSetHandleT, SectionAddrSet, ObjSetHandleCompare>
+      UnfinalizedSections;
+
+  std::vector<object::OwningBinary<object::Archive>> Archives;
+};
+}
+
+#endif // LLVM_LIB_EXECUTIONENGINE_ORC_MCJITREPLACEMENT_H
diff --git a/llvm/lib/ExecutionEngine/Orc/OrcTargetSupport.cpp b/llvm/lib/ExecutionEngine/Orc/OrcTargetSupport.cpp
new file mode 100644
index 00000000000..4a83a6c0b0c
--- /dev/null
+++ b/llvm/lib/ExecutionEngine/Orc/OrcTargetSupport.cpp
@@ -0,0 +1,102 @@
+#include "llvm/ADT/Triple.h"
+#include "llvm/ExecutionEngine/Orc/IndirectionUtils.h"
+
+#include <array>
+
+using namespace llvm;
+
+namespace {
+
+const char *JITCallbackFuncName = "call_jit_for_lazy_compile";
+const char *JITCallbackIndexLabelPrefix = "jit_resolve_";
+
+std::array<const char *, 12> X86GPRsToSave = {{
+    "rbp", "rbx", "r12", "r13", "r14", "r15", // Callee saved.
+    "rdi", "rsi", "rdx", "rcx", "r8", "r9",   // Int args.
+}};
+
+std::array<const char *, 8> X86XMMsToSave = {{
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" // FP args
+}};
+
+template <typename OStream> unsigned saveX86Regs(OStream &OS) {
+  for (const auto &GPR : X86GPRsToSave)
+    OS << "  pushq   %" << GPR << "\n";
+
+  OS << "  subq    $" << (16 * X86XMMsToSave.size()) << ", %rsp\n";
+
+  for (unsigned i = 0; i < X86XMMsToSave.size(); ++i)
+    OS << "  movdqu  %" << X86XMMsToSave[i] << ", "
+       << (16 * (X86XMMsToSave.size() - i - 1)) << "(%rsp)\n";
+
+  return (8 * X86GPRsToSave.size()) + (16 * X86XMMsToSave.size());
+}
+
+template <typename OStream> void restoreX86Regs(OStream &OS) {
+  for (unsigned i = 0; i < X86XMMsToSave.size(); ++i)
+    OS << "  movdqu  " << (16 * i) << "(%rsp), %"
+       << X86XMMsToSave[(X86XMMsToSave.size() - i - 1)] << "\n";
+  OS << "  addq    $" << (16 * X86XMMsToSave.size()) << ", %rsp\n";
+
+  for (unsigned i = 0; i < X86GPRsToSave.size(); ++i)
+    OS << "  popq    %" << X86GPRsToSave[X86GPRsToSave.size() - i - 1] << "\n";
+}
+
+uint64_t call_jit_for_fn(JITResolveCallbackHandler *J, uint64_t FuncIdx) {
+  return J->resolve(FuncIdx);
+}
+}
+
+namespace llvm {
+
+std::string getJITResolveCallbackIndexLabel(unsigned I) {
+  std::ostringstream LabelStream;
+  LabelStream << JITCallbackIndexLabelPrefix << I;
+  return LabelStream.str();
+}
+
+void insertX86CallbackAsm(Module &M, JITResolveCallbackHandler &J) {
+  uint64_t CallbackAddr =
+      static_cast<uint64_t>(reinterpret_cast<uintptr_t>(call_jit_for_fn));
+
+  std::ostringstream JITCallbackAsm;
+  Triple TT(M.getTargetTriple());
+
+  if (TT.getOS() == Triple::Darwin)
+    JITCallbackAsm << ".section __TEXT,__text,regular,pure_instructions\n"
+                   << ".align 4, 0x90\n";
+  else
+    JITCallbackAsm << ".text\n"
+                   << ".align 16, 0x90\n";
+
+  JITCallbackAsm << "jit_object_addr:\n"
+                 << "  .quad " << &J << "\n" << JITCallbackFuncName << ":\n";
+
+  uint64_t ReturnAddrOffset = saveX86Regs(JITCallbackAsm);
+
+  // Compute index, load object address, and call JIT.
+  JITCallbackAsm << "  movq    " << ReturnAddrOffset << "(%rsp), %rax\n"
+                 << "  leaq    (jit_indices_start+5)(%rip), %rbx\n"
+                 << "  subq    %rbx, %rax\n"
+                 << "  xorq    %rdx, %rdx\n"
+                 << "  movq    $5, %rbx\n"
+                 << "  divq    %rbx\n"
+                 << "  movq    %rax, %rsi\n"
+                 << "  leaq    jit_object_addr(%rip), %rdi\n"
+                 << "  movq    (%rdi), %rdi\n"
+                 << "  movabsq $" << CallbackAddr << ", %rax\n"
+                 << "  callq   *%rax\n"
+                 << "  movq    %rax, " << ReturnAddrOffset << "(%rsp)\n";
+
+  restoreX86Regs(JITCallbackAsm);
+
+  JITCallbackAsm << "  retq\n"
+                 << "jit_indices_start:\n";
+
+  for (JITResolveCallbackHandler::StubIndex I = 0; I < J.getNumFuncs(); ++I)
+    JITCallbackAsm << getJITResolveCallbackIndexLabel(I) << ":\n"
+                   << "  callq " << JITCallbackFuncName << "\n";
+
+  M.appendModuleInlineAsm(JITCallbackAsm.str());
+}
+}