main.py

import editdistance
import frontmatter
from hexdump2 import hexdump
import gradio as gr
import json
import shlex
import subprocess
import tempfile

from dist import levenshtein_with_wildcard, print_match_summary

description = frontmatter.load("README.md").content

def trim(str, n):
    return "\n".join(str.splitlines()[n:])

def trim_objdump(str):
    return trim(str, 7)

def disassemble_bytes(byte_data, architecture, options):
    with tempfile.NamedTemporaryFile(suffix=".bin", delete=False) as temp_bin_file:
        temp_bin_file.write(byte_data)
        temp_bin_file_name = temp_bin_file.name

    disassembly = subprocess.run(
        ["objdump", "-D", "-b", "binary", "-m", architecture, "-M", options, temp_bin_file_name],
        capture_output=True,
        text=True
    ).stdout
    disassembly = trim_objdump(disassembly)

    return disassembly

def compile(compiler, flags, source):
    # Create a temporary file for the C source code
    with tempfile.NamedTemporaryFile(suffix=".c", delete=False) as temp_c_file:
        temp_c_file.write(source.encode())
        temp_c_file_name = temp_c_file.name

    # Create a temporary file for the object file
    with tempfile.NamedTemporaryFile(suffix=".o", delete=False) as temp_o_file:
        temp_o_file_name = temp_o_file.name

    # Compile the C file to an object file
    result = subprocess.run(
        [compiler, "-c", temp_c_file_name]
        + shlex.split(flags)
        + ["-o", temp_o_file_name],
        capture_output=True,
        text=True,
    )
    compile_output = result.stdout + result.stderr

    # Create a temporary file for the raw bytes
    with tempfile.NamedTemporaryFile(suffix=".raw", delete=True) as raw_bytes_file:
        subprocess.run(
            [
                "objcopy",
                "--only-section",
                ".text",
                # XXX in reality we should probably look at the sections
                "--only-section",
                ".text.*",
                "-O",
                "binary",
                temp_o_file_name,
                raw_bytes_file.name,
            ]
        )
        compiled_bytes = raw_bytes_file.read()

    # Disassemble the object file
    disassembly = subprocess.run(
        ["objdump", "-dr", temp_o_file_name],
        capture_output=True,
        text=True
    ).stdout
    disassembly = trim_objdump(disassembly)

    # Relocs
    # relocs = subprocess.run(
    #     ["objdump", "-r", temp_o_file_name],
    #     capture_output=True,
    #     text=True
    # ).stdout
    # relocs = trim(relocs, 3)

    json_relocs = subprocess.run(
        ["llvm-readobj-19", "--elf-output-style=JSON", "--relocations", temp_o_file_name],
        capture_output=True,
        text=True,
    ).stdout
    json_relocs = json.loads(json_relocs)
    json_relocs = json_relocs[0]["Relocations"]
    json_relocs = [r["Relocation"] for d in json_relocs for r in d['Relocs']]
    # Filter out .text
    json_relocs = [r for r in json_relocs if r["Symbol"]["Name"] != ".text"]

    def reloc_type2size(s):
        match s:
            case "R_X86_64_PC32":
                return 4
            case "R_X86_64_PLT32":
                return 4
            case _:
                assert False, f"Unknown reloc {s}"

    relocs_byte_range = [range(r["Offset"], r["Offset"] + reloc_type2size(r["Type"]["Name"])) for r in json_relocs]
    # Flatten relocs_byte_range
    relocs_byte_range = [chr(i) for r in relocs_byte_range for i in r]


    print(f"relocs: {relocs_byte_range}")
    print(print_match_summary(b"test", compiled_bytes, wildcard_offsets_str2=relocs_byte_range))

    if result.returncode == 0:
        return json_relocs, compiled_bytes, compile_output, disassembly
    else:
        return None, None, compile_output, disassembly

def predict(target_bytes, source, compiler, flags, disasm_arch, disasm_options):
    target_bytes = bytes.fromhex(target_bytes)
    compiled_relocs, compiled_bytes, compile_output, compiled_disassembly = compile(compiler, flags, source)
    target_disassembly = disassemble_bytes(target_bytes, disasm_arch, disasm_options)

    if compiled_bytes is not None:
        return (
            hexdump(compiled_bytes, result="return"),
            hexdump(target_bytes, result="return"),
            editdistance.eval(compiled_bytes, target_bytes),
            compile_output,
            compiled_disassembly,
            compiled_relocs,
            target_disassembly
        )
    else:
        return (
            "Compilation failed",
            hexdump(target_bytes, result="return"),
            -1,
            compile_output,
            compiled_disassembly,
            compiled_relocs,
            target_disassembly
        )


def run():
    demo = gr.Interface(
        fn=predict,
        description=description,
        inputs=[
            gr.Textbox(
                lines=10,
                label="Bytes of Target Function (in hex)",
                value="b8 2a 00 00 00 c3",
            ),
            gr.Textbox(
                lines=10,
                label="Decompiled C Source Code",
                value="int x;\nint foo() { return x; }",
            ),
            gr.Textbox(label="Compiler", value="g++"),
            gr.Textbox(label="Compiler Flags", value="-O2"),
            gr.Textbox(label="Architecture (objdump -m)", value="i386"),
            gr.Textbox(label="Disassembler options (objdump -M)", value="x86-64")
        ],
        outputs=[
            gr.Textbox(label="Compiled bytes"),
            gr.Textbox(label="Target bytes"),
            gr.Number(label="Edit distance (lower is better)"),
            gr.Textbox(label="Compiler Output"),
            gr.Textbox(label="Compiled Disassembly"),
            gr.JSON(label="Compiled relocations", open=True),
            gr.Textbox(label="Target Disassembly"),
        ],
    )

    demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)


run()