concept-synthesizer

Automatically synthesizing C++20 template constraints for function templates

dependencies

This project was tested on MacOS (Apple M1), but it should also work on Linux platforms.

Requirements include python3, wget, unzip, git, cmake, make, clang++.

Building the synthesizer needs the source code of Clang/LLVM, which will be downloaded by the build script build.sh.

The error message reduction measurement script measure-error.py has two default measurement targets (STL algorithm and Boost special_functions), where the Boost library will be automatically downloaded.

build

./build.sh

run

run the synthesizer on a single C++ file and interpret the result

./llvm-project-llvmorg-19.1.1/build/bin/concept-synthesizer <c++-file> -- -std=c++20

You may need to tell the synthesizer include search paths for your C++ file, via -I. You can also use a pre-processed C++ file.

The synthesizer outputs five "sections" to stdout.

• Individual results: (nontrivial) synthesis information for each type template parameter
• Invalid calls: generated calls to these function templates (used to test error messages)
• Statistics: synthesis statistics
• Constrained code: this is the rewritten C++ code
• Resource consumption: currently only reports the time taken

For example, running the synthesizer on examples/demo.cpp outputs the following result (certain details are omitted).

[-[Individual results]-]
...
[-[]-]
[-[Invalid calls]-]
...
[-[]-]
[-[Statistics]-]
...
[-[]-]
[-[Constrained code]-]
template <typename T>
// added by concept-synth, original LN: 1
requires
requires (T o) { o.F(); }
void f(T x) {
    x.F();
}

template <typename U>
// added by concept-synth, original LN: 6
requires
requires (U o) { o.G(); }
void g(U x) {
    x.G();
}

template <typename V>
// added by concept-synth, original LN: 11
requires
(
 requires (V o) { o.G(); } &&
 requires (V o) { o.F(); } &&
 requires (V x0) { x0++; }
)
void h(V x) {
    x++;
    if (x) {
        f(x);
    } else {
        g(x);
    }
}

int main() {
}

[-[]-]
[-[Resource consumption]-]
...
[-[]-]

run the automatic error message reduction measurement script

This script assumes there is a command clang++ on the machine. So it is testing the error messages of that compiler.

python3 measure-error.py

docker images

It's recommended to build a linux/amd64 image which can then be run on both amd64 and Apple M1 (arm64) (through Rosetta 2) as described here.

The reverse direction looks more complicated.

Multi-platform images also look more complicated.

known issues

If you're experiencing compatibility issues (e.g. you use the clang++ on your machine to pre-process C++ files but then use this in-source built synthesizer on the pre-processed files), you can consider building clang and libcxx from the source as well.

It is known that the detailed experimental numbers may change among different runs. The reason is being investigated, but should be non-essential.