CWhy

by Emery Berger, Nicolas van Kempen, and Bryce Adelstein Lelbach.

("See why")

Explains and suggests fixes for compiler error messages for a wide range of programming languages, including C, C++, C#, Go, Java, LaTeX, PHP, Python, Ruby, Rust, Swift, and TypeScript.

Installation

[!NOTE]

CWhy needs to be connected to an OpenAI account. Your account will need to have a positive balance for this to work (check your OpenAI balance). Get an OpenAI key here.

You may need to purchase $0.50 - $1 in OpenAI credits depending on when your API account was created.

Once you have an API key, set it as an environment variable called OPENAI_API_KEY.

 # On Linux/MacOS:
 export OPENAI_API_KEY=<your-api-key>

 # On Windows:
 $env:OPENAI_API_KEY=<your-api-key>

python3 -m pip install cwhy

Other LLMs

We mostly test with OpenAI, but other LLMs can be made to work with CWhy. Please report any bug you may encounter.

OpenAI API Compatible

If your provider supports OpenAI style API calls, you can simply specify the OPENAI_BASE_URL environment variable to select a different URL to send requests to. For example, this will work great with Ollama:

docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --rm --name ollama ollama/ollama
docker exec -it ollama ollama pull llama3.1:70b
export OPENAI_BASE_URL=http://localhost:11434/v1
cwhy --llm llama3.1:70b --- clang++ tests/c++/missing-hash.cpp

LiteLLM Proxy

If your provider does not support OpenAI style API calls, such as AWS Bedrock which we used to support, we recommend using the LiteLLM Proxy Server.

pip install 'litellm[proxy]'
# Set AWS_ACCESS_KEY_ID, AWS_REGION_NAME, and AWS_SECRET_ACCESS_KEY.
litellm --model bedrock/anthropic.claude-v2
export OPENAI_BASE_URL=http://0.0.0.0:4000
cwhy --- clang++ tests/c++/missing-hash.cpp

Note that when using the LiteLLM Proxy, CWhy's --llm argument will be ignored completely.

Usage

Linux/MacOS

The wrapper mode is now default and mandatory, with a slightly modified interface. CWhy can either be used standalone by passing the full command after the triple dashes ---, or as part of a build tool by creating a short executable script wrapping the compiler command.

# Invoking the compiler directly.
cwhy --- g++ mycode.cpp

# Using CWhy with Java and an increased timeout.
cwhy --timeout 180 --- javac MyCode.java

# Invoking with GNU Make, using GPT-3.5.
CXX=`cwhy --llm=gpt-3.5-turbo --wrapper --- c++` make

# Invoking with CMake, using GPT-4 and clang++.
CWHY_DISABLE=1 cmake -DCMAKE_CXX_COMPILER=`cwhy --llm=gpt-4 --wrapper --- clang++` ...

Configuration tools such as CMake or Autoconf will occasionally invoke the compiler to check for features, which will fail and invoke CWhy unnecessarily if not available on the machine. To circumvent this, CWHY_DISABLE can be set in the environment to disable CWhy at configuration time.

CWHY_DISABLE='ON' cmake -DCMAKE_CXX_COMPILER=`cwhy --wrapper --- c++` ...

Windows

Windows support has been tested using Powershell. On the command line, using Ninja is required as MSBuild / .vcxproj will override any option set.

$env:CWHY_DISABLE='ON'
cmake -G Ninja -DCMAKE_CXX_COMPILER="$(python -m cwhy --wrapper --- cl)"  ...
$env:CWHY_DISABLE=''

Continuous Integration

CI using GitHub actions is straightforward on both Linux and MacOs. On Windows, Ninja is not installed by default on the image, and cl is not bound to the compiler. We recommend using choco install ninja and ilammy/msvc-dev-cmd to work around these two issues.

An example action YAML file covering all three platforms can be found here.

Important: Set the CWHY_DISABLE environment variable at configure-time to save money and cycles.

Options

These options can be displayed with cwhy --help.

• --llm: pick a specific OpenAI LLM. CWhy has been tested with gpt-3.5-turbo and gpt-4.
• --timeout: pick a different timeout than the default for API calls.
• --show-prompt (debug): print prompts before calling the API.

Examples

C++

This highlighted example is missing-hash.cpp, which is one of the first cases we experimented with.

And here's the English-language explanation from cwhy:

% cwhy --- clang++ --std=c++20 -c missing-hash.cpp
The problem is that you are attempting to use `std::unordered_set`
with `std::pair<int, int>` as the key type. However, the standard
library does not provide a hash function specialization for
`std::pair` out of the box, so the default constructor of the
unordered set is deleted.

To resolve this, you'll need to provide a custom hash function for
`std::pair<int, int>`. Here's an example of how you can define one:

```cpp
struct PairHash {
    template <typename T1, typename T2>
    std::size_t operator()(const std::pair<T1, T2>& pair) const {
        std::hash<T1> hash1;
        std::hash<T2> hash2;
        return hash1(pair.first) ^ (hash2(pair.second) << 1);
    }
};

Then, when instantiating the std::unordered_set, you can specify the custom hash function:

std::unordered_set<std::pair<int, int>, PairHash> visited;

With this change, the code should now compile and work as expected.

### Rust

% cwhy --- cargo build There are three issues: 1. There are two unused variables x in the code. 2. The variable x is used after it has already been moved in the call to f(x), which takes ownership of x. 3. The function f(x) takes ownership of x, which may not be necessary and could be changed to borrow the value instead.