This artifact is intended to reproduce the experimental results presented in our paper, "Towards Sound Reassembly of Modern x86-64 Binaries", published at ASPLOS '25. It contains scripts for running experiments and datasets we used.

This artifact will answer all three research questions from our paper:

• RQ1: How well does SURI compare to the state-of-the-art reassembly tools in terms of reliability?
• RQ2: How big is the performance overhead introduced by SURI for rewritten binaries?
• RQ3: Is SURI applicable to real-world scenarios, such as runtime memory sanitization?

To answer these questions, we conducted total 5 experiments as follows:

• Exp1: Reassembly completion comparison (RQ1)
• Exp2: Test suite pass rate comparison (RQ1)
• Exp3: Reliability test on real-world programs (RQ1)
• Exp4: Reassembly overhead measurement (RQ2)
• Exp5: Application of SURI (RQ3)

We have three comparison targets for the comparative study of SURI.

• Ddisasm: a binary reassembler based on datalog disassembly (USENIX Security '20)
• Egalito: a binary recompiler based on layout-agnostifc binary recompilation (ASPLOS '20)
• BASan: a binary-only address sanitizer implemented on top of RetroWrite (S&P '20)

This table is a brief summary of each tool:

We used 5 different kinds of benchmark programs to evaluate SURI:

• Coreutils v9.1
• Binutils v2.40
• SPEC CPU 2006 v1.2 and 2017 v1.1.5
• 10 real-world programs
  • Apache v2.4.56
  • MariaDB v11.5.0
  • Nginx v1.23.3
  • SQLitev 3.31.2
  • 7-Zip-24.05
  • Epiphany-3.36.4
  • Filezilla v3.46.3
  • Openssh v8.2p1
  • Putty v0.73
  • Vim v8.1
• Juliet Test Suite v1.3

Coreutils, Binutils, and SPEC are used for Exp1, Exp2, and Exp4, real-world programs are used for Exp3, and Juliet Test Suite is used for Exp5.

For Coreutils, Binutils and SPEC benchmarks, we further make three different datasets:

• setA: binaries compiled on Ubuntu 20.04 (SURI vs. Ddisasm)
• setB: binaries compiled on Ubuntu 18.04 (SURI vs. Egalito)
• setC: binaries compiled on Ubuntu 20.04 w/o call frame information (ablation study - see Section 4.3.3 of the paper)

Below table sumarizes our datasets:

⚠️ We exclude SPEC benchmark binaries from our dataset because they are proprietary. However, we prepare benchmark building scripts for SPEC benchmarks, in case you have a valid license of SPEC CPU 2006 or 2017. Our experimental scripts will work well regardless of the existence of SPEC binaries, though.

These are the links that explain how to set up our artifact and how to run the experiments.

• Preparation
• Experiment