LLM4Decompile Framework

Transform raw pseudo-code into normalized representations for training using the following steps in the Preprocess/ directory.

1. Install Preprocessing Requirements:

cd Preprocess
pip install tree-sitter==0.24.0 tree-sitter-c==0.23.4 tqdm

2. Execute Preprocessing Pipeline:

• Normalize pseudo-code (R2I standard): python3 normalize_pseudo.py --input_json <input.json> --output_json <output.json> --key_name pseudo
• Obfuscate source code to generate IR: python3 normalize_src_basedonpseudo.py --input_json <input.json> --output_json <output.json> --top 0 --pseudo pseudo_norm
• Format codes with clang-format: python3 format.py --input <input.json> --output <output.json>
• Infer types for obfuscated IR: python3 inf_type.py --input_json <input.json> --output_name <output_name> --generator ../psychec/psychecgen --solver ../psychec/psychecsolver-exe --split 2 --idx 0

cd Preprocess
# Requirements
pip install tree-sitter==0.24.0 tree-sitter-c==0.23.4 tqdm

# Step 1: Normalize pseudo-code according to R2I standard
python3 normalize_pseudo.py --input_json exebench_c.json --output_json exebench_pseudonorm.json --key_name pseudo

# Step 2: Obfuscate source code to generate IR
python3 normalize_src_basedonpseudo.py --input_json exebench_pseudonorm.json --output_json exebench_norm_top0.json --top 0 --pseudo pseudo_norm

# Step 3: Format codes with clang-format
python3 format.py --input exebench_norm_top0.json --output exebench_format_top0.json

# Step 4: Infer types for obfuscated IR (used for compiler-based rewards)
python3 inf_type.py --input_json train_format_top0.json --output_name train_format_top0_type \
    --generator ../psychec/psychecgen --solver ../psychec/psychecsolver-exe --split 2 --idx 0

Use Ghidra's headless analyzer to decompile a binary into pseudo-code. This process involves compiling C code into a binary, then running the headless analyzer with a post-script to extract the function's pseudo-code.

Note: You must replace func0 in the script with the actual name of the function you wish to decompile.

Required variables to configure in the script:

• ghidra_path: Path to the headless analyzer (e.g., ./ghidra_11.0.3_PUBLIC/support/analyzeHeadless).
• postscript: Path to your decompiler helper function (e.g., ./decompile.py).
• project_path: Temporary folder for analysis.
• func_path: Path to the C source code for compilation.
• fileName: Base name for output files.

import os
import subprocess
from tqdm import tqdm,trange

OPT = ["O0", "O1", "O2", "O3"]
timeout_duration = 10

ghidra_path = "./ghidra_11.0.3_PUBLIC/support/analyzeHeadless"
postscript = "./decompile.py"
project_path = "."
project_name = "tmp_ghidra_proj"
func_path = "../samples/sample.c"
fileName = "sample"

# ... (rest of the subprocess and Ghidra execution logic) ...
        command = [
            ghidra_path,
            temp_dir,
            project_name,
            "-import", executable_path,
            "-postScript", postscript, output_path,
            "-deleteProject",
        ]
        result = subprocess.run(command, text=True, capture_output=True, check=True)

If you only want to reproduce the evaluation step (Step 5) using pre-built data, follow these steps. You must have the Bringup-Bench source repository cloned locally.

1. Clone the benchmark repository:

   git clone https://github.com/toddmaustin/bringup-bench.git

2. Configure the environment paths in config.env by setting BENCH_REPO_ROOT to your local bringup-bench path.
3. Run the evaluation script (e.g., for optimization level O0):

   python3 scripts/eval_infer_out.py data/infer_results/merged.O0.func_map.infer.jsonl

4. View the results in the reports/ directory.

# 1. Clone Bringup-Bench
git clone https://github.com/toddmaustin/bringup-bench.git

# 2. Configure paths
cd bringupbench
# Set BENCH_REPO_ROOT to your bringup-bench path in config.env

# 3. Run evaluation (e.g., O0)
python3 scripts/eval_infer_out.py data/infer_results/merged.O0.func_map.infer.jsonl

# 4. Check results
cat reports/O0_results.md