Skip to content
c++ hot code reload for linux and macos
Branch: master
Clone or download
Latest commit d16bb65 May 17, 2019

README.md

Demo

Build Status

jet-live

jet-live is a library for c++ "hot code reloading". It works on linux and modern macOS (10.12+ I guess) on 64 bit systems powered by cpu with x86-64 instruction set. Apart from reloading of functions it is able to keep apps' static and global state unchanged after code was reload (please refer to "How it works" for what is it and why it is important). Tested on Ubuntu 18.04 with clang 6.0.1/7.0.1, lld-7, gcc 6.4.0/7.3.0, GNU ld 2.30, cmake 3.10.2, ninja 1.8.2, make 4.1 and macOS 10.13.6 with Xcode 8.3.3, cmake 3.8.2, make 3.81.

Important: this library doesn't force you to organize your code in some special way (like in RCCPP or cr), you don't need to separate reloadable code into some shared library, jet-live should work with any project in the least intrusive way.

If you need something similar for windows, please try blink, I have no plans to support windows.

Prerequisites

You need c++11 compliant compiler. Also there're several dependencies which are bundled in, most of them are header-only or single h/cpp pair library. Please refer to the lib directory for details.

Getting started

This library is best suited for projects based on cmake and make or ninja build systems, defaults are fine-tuned for these tools. The CMakeLists.txt will add set(CMAKE_EXPORT_COMPILE_COMMANDS ON) option for compile_commands.json and alter compiler and linker flags. This is important and not avoidable. For details please see CMakeLists.txt. if you use ninja, add -d keepdepfile ninja flag when running ninja, this is needed to track dependencies between source and header files

  1. In your project CMakeLists.txt file:
include(path/to/jet-live/cmake/jet_live_setup.cmake) # setup needed compiler and linker flags, include this file in your root CMakeLists.txt
set(JET_LIVE_BUILD_EXAMPLE OFF)
set(JET_LIVE_SHARED ON) # if you want to
add_subdirectory(path/to/jet-live)
target_link_libraries(your-app-target jet-live)
  1. Create an instance of jet::Live class
  2. In your app runloop call liveInstance->update()
  3. When you need to reload code, call liveInstance->tryReload()

Important: This library is not thread safe. It uses threads under the hood to run compiler, but you should call all library methods from the same thread.

Also I use this library only with debug builds (-O0, not stripped, without -fvisibility=hidden and things like that) to not deal with optimized and inlined functions and variables. I don't know how it works on highly optimized stripped builds, most likely it will not work at all.

Personally I use it like this. I have a Ctrl+r shortcut to which tryReload is assigned in my application. Also my app app calls update in the main runloop and listens for onCodePreLoad and onCodePostLoad events to recreate some objects or re-evaluate some functions:

  1. I start my application
  2. I edit some files, save it, and now I know that I'm ready to reload new code (here previously I recompiled application)
  3. I press Ctrl+r
  4. I'm watching for output of my application to see if there're any compilation/linkage errors (see "Customizations")
  5. Both for success or fail, go to 2

jet-live will monitor for file changes, recompile changed files and only when tryReload is called it will wait for all current compilation processes to finish and reload new code. Please don't call tryReload on each update, it will not work as you're expecting, call it only when your source code is ready to be reloaded.

If you don't want to switch back and forth between your code editor and app, you can configure a keyboard shortcut which runs a shell command kill -s USR1 $(pgrep <your_app_name>), the library will trigger code reload when SIGUSR1 signal is received. It works at least in emacs, Xcode, CLion and VSCode, but I'm sure it is achievable in other editors and IDEs, just google it. If your debugger is lldb and it catches this signal and stops the app, add this commands to the ~/.lldbinit file:

breakpoint set --name main
breakpoint command add
process handle -n true -p true -s false SIGUSR1
continue
DONE

On macOS you can use cmake -G Xcode generator apart from make and ninja. In this case please install xcpretty gem:

gem install xcpretty

Example

There's a simple example app, just run:

git clone https://github.com/ddovod/jet-live.git && cd jet-live
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Debug .. && make
./example/example

and try hello command. Don't forget to run reload command after fixing the function.

Tests

There's a not very comprehensive, but constantly updating test suite. To run it:

git clone https://github.com/ddovod/jet-live.git && cd jet-live
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Debug -DJET_LIVE_BUILD_TESTS=ON .. && make
../tools/tests/test_runner.py -b . -s ../tests/src/

Features

Implemented:

  • Reloading functions
  • Relocating static and global state
  • Tracking dependencies
  • Working with code from this executable and loaded shared libraries
  • Linux and macOS implementation
  • Ability to add new compilations units on the fly (just invoke cmake to recreate compile_commands.json file after new .cpp file was created)

Will be implemented:

  • Code reload in multithreaded app (right now reloading of code in multithreaded app is not reliable)

Will not be implemented at all:

  • Reliable reload of lambda functions with non-empty captures. Lambdas with empty capture list are ok since they are just a plain functions at the lowest level (at least they are implemented this way). There's only 1 case where we can handle lambdas with non-empty capture list properly - if old and new code has lambda with exactly same signature and exactly same lambdas before this lambda within this file, in other cases the reload of lambdas is not reliable. The reason for this is mangled name of lambda type depends on the arguments and position of this lambda relative to another lambdas in this compilation unit. Moreover different compilers use different name mangling of lambda types. Please refer to tests to see good and bad cases.

Customizations

jet-live is fine-tuned to work with cmake and make/ninja tools, but if you want to adopt it to another build tool, there's a way to customize its' behaviour in some aspects. Please refer to sources and documentation. Also it is a good idea to create your own listener to receive events from the library. Please refer to documentation of ILiveListener and LiveConfig.

Important: it is highly recommended to log all messages from the library using ILiveListener::onLog to see if something went wrong.

How it works (for curious ones)

The library reads elf headers and sections of this executable and all loaded shared libraries, finds all symbols and tries to find out which of them can either be hooked (functions) or should be transferred/relocated (static/global variables). Also it finds symbols size and "real" address.

Apart from that jet-live tries to find compile_commands.json near your executable or in its' parent directories recursively. Using this file it distinguishes:

  • compilation command
  • source file path
  • .o (object) file path
  • .d (depfile) files path
  • compiler path
  • working directory - directory from which the compiler was run
  • some compiler flags for further processing.

When all compilation units are parsed, it distinguishes the most common directory for all source files and starts watching for all directories with source files, their dependencies and some service files like compile_commands.json.

Apart from that the library tries to find all dependencies for each compilation unit. By default it will read depfiles near the object files (see -MD compiler option). Suppose the object file is located at:

/home/coolhazker/projects/some_project/build/main.cpp.o

jet-live will try to find depfile at:

/home/coolhazker/projects/some_project/build/main.cpp.o.d
or
/home/coolhazker/projects/some_project/build/main.cpp.d

It will pick up all dependencies which are under the watching directories, so things like /usr/include/elf.h will not be treated as dependency even if this file is really included in some of your .cpp files.

Now the library is initialized.

Next, when you edit some source file and save it, jet-live immediately starts compilation of all dependent files in the background. By default the number of simultaneous compilation processes is 4, but you can configure it. It will write to log about successes and errors using ILiveListener::onLog method of listener. If you trigger compilation of some file when it is already compiling (or waiting in the queue), old compilation process will be killed and new one will be added to the queue, so its kinda safe to not wait for compilation to finish and make new changes of the code. Also after each file was compiled, it will update dependencies for compiled file since compiler can recreate depfile for it if new version of compilation unit has new dependencies.

When you call Live::tryReload, the library will wait for unfinished compilation processes and then all accumulated new object files will be linked together in shared library and placed near your executable with name lib_reloadXXX.so, where XXX is a number of "reloads" during this session. So lib_reloadXXX.so contains all new code.

jet-live loads this library using dlopen, reads elf/mach-o headers and sections and finds all symbols. Also it loads relocation info from the object files which was used to construct this new library. After that:

  • For all hookable functions it transfers the control flow from old version to new one
  • For all link-time relocations it fixes new shared library in a way where new code is pointing to old already living static/global variables
  • For all local static variables that were not relocated it just memcpy memory from old location to new one

Important: ILiveListener::onCodePreLoad event is fired right before lib_reloadXXX.so is loaded into the process memory. ILiveListener::onCodePostLoad event is fired right after all code-reloading-machinery is finished.

About functions hooking

You can read more about function hooking here. This library uses awesome subhook library to redirect function flow from old to new ones. You can see that on 32 bit platforms your functions should be at least 5 bytes long to be hookable. On 64 bit you need at least 14 bytes which is a lot, and for example empty stub function will probably not fit into 14 bytes. From my observations, clang by default produces code with 16-byte functions alignment. GCC don't do this by default, so for GCC the -falign-functions=16 flag is used. That means the spacing between begins of any 2 functions is not less that 16 bytes, which makes possible to hook any function.

About state transfer

New versions of functions should use statics and globals which are already living in the application. Why is it important? Suppose you have (a bit synthetic example, but anyway):

// Singleton.hpp
class Singleton
{
public:
    static Singleton& instance();
};

int veryUsefulFunction(int value);

// Singleton.cpp
Singleton& Singleton::instance()
{
    static Singleton ins;
    return ins;
}

int veryUsefulFunction(int value)
{
    return value * 2;
}

Then you want to update veryUsefulFunction to smth like this:

int veryUsefulFunction(int value)
{
    return value * 3;
}

Great, now it multiplies argument by 3. But since whole Singleton.cpp will be reloaded and Singleton::instance function will be hooked to call new version, lib_reloadXXX.so will contain new static variable static Singleton ins, which is not initialized, and if you call Singleton::instance() after code was reloaded, it will initialize this variable again which is not good cause we don't want to call its constructor again. Thats why we need to relocate all statics and globals to the new code and transfer the guard variables of statics. Most of the link-time relocations related to statics and globals are 32-bit. So if the shared library with new code will be loaded too far in memory from the application, it will be not possible to relocate variables in this way. To solve this, new shared library is linked using special linker flags which allows us to load it into specific pre-calculated location in the virtual memory (see -image_base in Apple ld, --image-base in LLVM lld and -Ttext-segment + -z max-page-size in GNU ld linker flags).

Also your app will probably crash if you try to change memory layout of your data types in reloadable code.

Suppose you have an instance of this class allocated somewhere in the heap or on the stack:

class SomeClass
{
public:
    void calledEachUpdate() {
        m_someVar1++;
    }
private:
    int m_someVar1 = 0;
}

You edit it and now it looks like:

class SomeClass
{
public:
    void calledEachUpdate() {
        m_someVar1++;
        m_someVar2++;
    }
private:
    int m_someVar1 = 0;
    int m_someVar2 = 0;
}

After code is reloaded, you'll probably observe a crash because already allocated object has different data layout, it has no m_someVar2 instance variable, but new version of calledEachUpdate will try to modify it actually modifying random data. In such cases you should delete this instance in onCodePreLoad callback and recreate it in onCodePostLoad callback. Correct transfer of its state is up to you. The same effect will take place if you'll try to change static data structures layout. The same also correct for polymorphic classes (vtable) and lambdas with captures (captures are stored inside lambdas' data fields).

Licence

MIT

For licences of used libraries please refer to their directories and source code.

You can’t perform that action at this time.