Please follow the command to prepare dependency.
conda create -n rebench python=3.12
conda activate rebench
pip install -r requirements.txt
REBench can be generated via two decmopilers, ghidra and IDA Pro. We utilized Ghidra 11.4.2 and IDA 7.6. Ghidra decompiler requires preprocessing.
/path/to/ghidra/support/analyzeHeadless /proj-dir /proj-name -import ./binary -readOnly -postScript ./ghidra_preprocess_1.py
/path/to/ghidra/support/analyzeHeadless /proj-dir /proj-name -import ./strip_binary -readOnly -postScript ./ghidra_postprocess_1.py
The binary should contain symbols and the second script should execute with stripped binaries.
IDA does not need preprocessing.
/path/to/ida/idat -L"log" -A -S"./ida_process.py" binary
We provide finetuning and running scripts for CodeLlama.
python finetune_codellma.py
python codellama_script.py --arch arch --opt opt
To evaluate the name recovery, please follow the execution steps from the name_eval directory.
python parse.py codellama x64 O0
python evaluation.py --evaluation-input evaluation_input_codellama_x64_O0_func.txt
python evaluation.py --evaluation-input evaluation_input_codellama_x64_O0_var.txt
To evaluate the type inference, please follow the execution step from the type_eval directory.
python type_evaluation.py --model model_name --arch arch --opt opt
There are binaries and decompiled code in the following link.