JitFromScratch

Collection of examples from my talks in the LLVM Social Berlin and C++ User Group Berlin that implement various aspects of a JIT compiler based on the LLVM Orc libraries.

Contribute

The repository follows a perfect history policy to foster traceability and understanding. If you find a bug or want to submit an improvement, please file a pull request against the respective step.

Structure

The examples are organized around a nonsense command line program that compiles the code for a simple function at runtime:

template <size_t sizeOfArray>
int *integerDistances(const int (&x)[sizeOfArray], int *y) {
  int items = arrayElements(x);
  int *results = customIntAllocator(items);

  for (int i = 0; i < items; i++) {
    results[i] = abs(x[i] - y[i]);
  }

  return results;
}

Steps

The example project is built in a series of self-contained steps. You can immediatedly explore and build each change in your Browser with GitPod:

Step	Change	Travis	GitPod	Description
0	87cb540			Wire up LLVM
1	82465fc			Add lit tests
2	b1b268b			Add TravisCI config to run tests in prebuilt docker
3	81a9d8b			Initialize LLVM for native target codegen
4	731f482			Add minimal JIT compiler based on LLJIT
5	73a565a			Create empty module and pass it to the JIT
6	7690ac1			In debug mode dump extra info when passing -debug -debug-only=jitfromscratch
7	3b47ce8			Generate function that takes two ints and returns zero
8	8a4bba8			Add basic sanity checks for IR code
9	402e9d1			Request and run trivial function
10	e08ff57			Emit IR code for substraction and use it in the integerDistances function
11	8cf690b			Emit variable names to make IR more readable
12	03b8701			Allow symbol resolution from the host process
13	a0cdaad			Emit call to stdlib function abs in JITed code
14	2653289			Emit unrolled loop in JITed code
15	2291ee4			Emit call to allocator in JITed code
16	bf72028			Remove wrapper function and call JITed code directly
17	96f4c8e			Break free from LLJIT
18	a4b88a0			Implement GDB JIT interface
19	b7212c3			Add optimization passes controlled via -Ox command line flag
20	75dff62			Simple name-based object cache

Build and run locally

$ git clone https://github.com/weliveindetail/JitFromScratch jitfromscratch
$ mkdir build && cd build
$ cmake -GNinja -DLLVM_DIR=/path/to/llvm-build/lib/cmake/llvm ../jitfromscratch
$ ninja run

The project was tested on Ubuntu 19.04 and macOS 10.14. Please find real-world examples for commands and expected output on Travis CI.

Build and run in docker

$ git clone https://github.com/weliveindetail/JitFromScratch jitfromscratch
$ cd jitfromscratch
$ docker run -it -v $(pwd):/host/jitfromscratch -e REMOTE=/host/jitfromscratch -e CHECKOUT=$(git rev-parse HEAD) weliveindetail/jitfromscratch:llvm09-test-release

Further images can be pulled from dockerhub or created manually based on the given Dockerfiles.

Setup vscode

Install the C/C++, CMake Tools and CodeLLDB extensions and use the given configuration files.

Previous Versions

You can easily diff for a specific 8.0 to 9.0 change set, e.g.:

$ git diff origin/llvm08/master origin/llvm09/master
$ git diff origin/llvm08/steps/A00 origin/llvm09/steps/A00 -- CMakeLists.txt
$ git diff origin/llvm08/steps/A18 origin/llvm09/steps/A18 -- JitFromScratch.cpp

In LLVM 8, the Orc library was almost entirely rewritten and I used the opportunity to also refine my examples. You can still find the old examples here:

• LLVM 8.0
• LLVM 5.0
• LLVM 4.0