4. Building a JIT: Extreme Laziness - Using Compile Callbacks to JIT from ASTs

This tutorial is under active development. It is incomplete and details may change frequently. Nonetheless we invite you to try it out as it stands, and we welcome any feedback.

4.1. Chapter 4 Introduction

Welcome to Chapter 4 of the “Building an ORC-based JIT in LLVM” tutorial. This chapter introduces the Compile Callbacks and Indirect Stubs APIs and shows how they can be used to replace the CompileOnDemand layer from Chapter 3 with a custom lazy-JITing scheme that JITs directly from Kaleidoscope ASTs.

To be done:

(1) Describe the drawbacks of JITing from IR (have to compile to IR first, which reduces the benefits of laziness).

(2) Describe CompileCallbackManagers and IndirectStubManagers in detail.

(3) Run through the implementation of addFunctionAST.

4.2. Full Code Listing

Here is the complete code listing for our running example that JITs lazily from Kaleidoscope ASTS. To build this example, use:

# Compile
clang++ -g toy.cpp `llvm-config --cxxflags --ldflags --system-libs --libs core orc native` -O3 -o toy
# Run

Here is the code:

//===----- KaleidoscopeJIT.h - A simple JIT for Kaleidoscope ----*- C++ -*-===//
//                     The LLVM Compiler Infrastructure
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
// Contains a simple JIT definition for use in the kaleidoscope tutorials.


#include "llvm/ADT/STLExtras.h"
#include "llvm/ExecutionEngine/ExecutionEngine.h"
#include "llvm/ExecutionEngine/JITSymbol.h"
#include "llvm/ExecutionEngine/RuntimeDyld.h"
#include "llvm/ExecutionEngine/SectionMemoryManager.h"
#include "llvm/ExecutionEngine/Orc/CompileOnDemandLayer.h"
#include "llvm/ExecutionEngine/Orc/CompileUtils.h"
#include "llvm/ExecutionEngine/Orc/IRCompileLayer.h"
#include "llvm/ExecutionEngine/Orc/IRTransformLayer.h"
#include "llvm/ExecutionEngine/Orc/LambdaResolver.h"
#include "llvm/ExecutionEngine/Orc/ObjectLinkingLayer.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/LegacyPassManager.h"
#include "llvm/IR/Mangler.h"
#include "llvm/Support/DynamicLibrary.h"
#include "llvm/Support/Error.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Scalar/GVN.h"
#include <algorithm>
#include <cassert>
#include <cstdlib>
#include <memory>
#include <string>
#include <vector>

class PrototypeAST;
class ExprAST;

/// FunctionAST - This class represents a function definition itself.
class FunctionAST {
  std::unique_ptr<PrototypeAST> Proto;
  std::unique_ptr<ExprAST> Body;

  FunctionAST(std::unique_ptr<PrototypeAST> Proto,
              std::unique_ptr<ExprAST> Body)
      : Proto(std::move(Proto)), Body(std::move(Body)) {}

  const PrototypeAST& getProto() const;
  const std::string& getName() const;
  llvm::Function *codegen();

/// This will compile FnAST to IR, rename the function to add the given
/// suffix (needed to prevent a name-clash with the function's stub),
/// and then take ownership of the module that the function was compiled
/// into.
irgenAndTakeOwnership(FunctionAST &FnAST, const std::string &Suffix);

namespace llvm {
namespace orc {

class KaleidoscopeJIT {
  std::unique_ptr<TargetMachine> TM;
  const DataLayout DL;
  ObjectLinkingLayer<> ObjectLayer;
  IRCompileLayer<decltype(ObjectLayer)> CompileLayer;

  typedef std::function<std::unique_ptr<Module>(std::unique_ptr<Module>)>

  IRTransformLayer<decltype(CompileLayer), OptimizeFunction> OptimizeLayer;

  std::unique_ptr<JITCompileCallbackManager> CompileCallbackMgr;
  std::unique_ptr<IndirectStubsManager> IndirectStubsMgr;

  typedef decltype(OptimizeLayer)::ModuleSetHandleT ModuleHandle;

      : TM(EngineBuilder().selectTarget()),
        CompileLayer(ObjectLayer, SimpleCompiler(*TM)),
                      [this](std::unique_ptr<Module> M) {
                        return optimizeModule(std::move(M));
            orc::createLocalCompileCallbackManager(TM->getTargetTriple(), 0)) {
    auto IndirectStubsMgrBuilder =
    IndirectStubsMgr = IndirectStubsMgrBuilder();

  TargetMachine &getTargetMachine() { return *TM; }

  ModuleHandle addModule(std::unique_ptr<Module> M) {

    // Build our symbol resolver:
    // Lambda 1: Look back into the JIT itself to find symbols that are part of
    //           the same "logical dylib".
    // Lambda 2: Search for external symbols in the host process.
    auto Resolver = createLambdaResolver(
        [&](const std::string &Name) {
          if (auto Sym = IndirectStubsMgr->findStub(Name, false))
            return Sym;
          if (auto Sym = OptimizeLayer.findSymbol(Name, false))
            return Sym;
          return JITSymbol(nullptr);
        [](const std::string &Name) {
          if (auto SymAddr =
            return JITSymbol(SymAddr, JITSymbolFlags::Exported);
          return JITSymbol(nullptr);

    // Build a singleton module set to hold our module.
    std::vector<std::unique_ptr<Module>> Ms;

    // Add the set to the JIT with the resolver we created above and a newly
    // created SectionMemoryManager.
    return OptimizeLayer.addModuleSet(std::move(Ms),

  Error addFunctionAST(std::unique_ptr<FunctionAST> FnAST) {
    // Create a CompileCallback - this is the re-entry point into the compiler
    // for functions that haven't been compiled yet.
    auto CCInfo = CompileCallbackMgr->getCompileCallback();

    // Create an indirect stub. This serves as the functions "canonical
    // definition" - an unchanging (constant address) entry point to the
    // function implementation.
    // Initially we point the stub's function-pointer at the compile callback
    // that we just created. In the compile action for the callback (see below)
    // we will update the stub's function pointer to point at the function
    // implementation that we just implemented.
    if (auto Err = IndirectStubsMgr->createStub(mangle(FnAST->getName()),
      return Err;

    // Move ownership of FnAST to a shared pointer - C++11 lambdas don't support
    // capture-by-move, which is be required for unique_ptr.
    auto SharedFnAST = std::shared_ptr<FunctionAST>(std::move(FnAST));

    // Set the action to compile our AST. This lambda will be run if/when
    // execution hits the compile callback (via the stub).
    // The steps to compile are:
    // (1) IRGen the function.
    // (2) Add the IR module to the JIT to make it executable like any other
    //     module.
    // (3) Use findSymbol to get the address of the compiled function.
    // (4) Update the stub pointer to point at the implementation so that
    ///    subsequent calls go directly to it and bypass the compiler.
    // (5) Return the address of the implementation: this lambda will actually
    //     be run inside an attempted call to the function, and we need to
    //     continue on to the implementation to complete the attempted call.
    //     The JIT runtime (the resolver block) will use the return address of
    //     this function as the address to continue at once it has reset the
    //     CPU state to what it was immediately before the call.
      [this, SharedFnAST]() {
        auto M = irgenAndTakeOwnership(*SharedFnAST, "$impl");
        auto Sym = findSymbol(SharedFnAST->getName() + "$impl");
        assert(Sym && "Couldn't find compiled function?");
        JITTargetAddress SymAddr = Sym.getAddress();
        if (auto Err =
                                              SymAddr)) {
          logAllUnhandledErrors(std::move(Err), errs(),
                                "Error updating function pointer: ");

        return SymAddr;

    return Error::success();

  JITSymbol findSymbol(const std::string Name) {
    return OptimizeLayer.findSymbol(mangle(Name), true);

  void removeModule(ModuleHandle H) {

  std::string mangle(const std::string &Name) {
    std::string MangledName;
    raw_string_ostream MangledNameStream(MangledName);
    Mangler::getNameWithPrefix(MangledNameStream, Name, DL);
    return MangledNameStream.str();

  std::unique_ptr<Module> optimizeModule(std::unique_ptr<Module> M) {
    // Create a function pass manager.
    auto FPM = llvm::make_unique<legacy::FunctionPassManager>(M.get());

    // Add some optimizations.

    // Run the optimizations over all functions in the module being added to
    // the JIT.
    for (auto &F : *M)

    return M;

} // end namespace orc
} // end namespace llvm


Next: Remote-JITing – Process-isolation and laziness-at-a-distance