Awesome

goSAT

goSAT is an SMT solver for the theory of floating-point arithmetic (FPA) that is based on global optimization.

Overview

goSAT is an SMT solver based on mathematical global optimization. It casts the satisfiability problem of FPA to a corresponding global optimization problem. It builds on the ideas proposed in XSat. Compared to XSat, we implemented the following key features:

• JIT compilation of SMT formulas using LLVM.
• Integration with NLopt, a feature-rich mathematical optimization package.

In our experiments, goSAT demonstrated better efficiency compared to z3 and mathsat by an order of magnitude or more. More details are available in our FMCAD'17 paper and this appendix.

Citing

If you use goSAT in an academic work, please cite:

@inproceedings{BenKhadra2017a,
author = {{Ben Khadra}, M Ammar and Stoffel, Dominik and Kunz, Wolfgang},
booktitle = {Proceedings of Formal Methods in Computer-Aided Design (FMCAD'17)},
doi = {10.23919/FMCAD.2017.8102235},
pages = {11--14},
title = {{goSAT: Floating-point Satisfiability as Global Optimization}},
year = {2017}
}

Dependencies

This project depends on:

• Z3 library for input file parsing and model validation
• LLVM for JIT execution of SMT formulas.
• NLopt for finding the minima of the objective function.

Installing z3 and nlopt should be straightforward. As for installing LLVM, you might find this tutorial to be useful.

goSAT is known to work with z3 v4.6, LLVM v4.0.1, nlopt v2.4.2. We recommend building LLVM from source as the official pre-built package might be broken. Specifically, we are aware of such issues on Ubuntu 16.04.

Building

You can build the project using a command like,

mkdir build; cd build
cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_PREFIX_PATH=/home/gosat/local/ -DLLVM_DIR=/home/gosat/local/llvm/lib/cmake/llvm/ ..
make

assuming that libnlopt and libz3 are installed at the prefix ${HOME}/local/ while llvm is installed at ${HOME}/local/llvm.

Usage

An SMT file needs to be provided using -f option. Additionally, the tool operation mode can be set. goSAT supports three operation modes:

• Native solving. This is the default mode where a given formula is first transformed to an objective function in LLVM IR. Then, the objective function is jitted using MCJIT and solved using the NLopt backend. This mode can be explicitly set using -mode=go option.

• Code generation. This mode can be enabled using -mode=cg option. It mimics the behavior of XSat where C code gets generated representing the objective function. Generated code needs to be compiled and solved using on of the tools we provide in tools folder. Moreover, you can specify an API dump format like -mode=cg -fmt=plain for generating several objective functions. The latter is useful for building a library as discussed here.

• Formula analysis. A simple analysis to show the number of variables, their types, and other misc facts about a given SMT formula. This mode is enabled using -mode=fa option.

The default output of goSAT is in csv format. It lists the benchmark name, sat result, elapsed time (seconds), minimum found, and status code returned by nlopt. The minimum found should be zero in case of sat. Use option -smtlib-output to obtain conventional solver output which is more succinct.

Note that goSAT relies on stochastic search which means that we can not generally prove an instance to be unsat even if it is actually so. Therefore, goSAT reports back either sat or unknown. Being stochastic, gives goSAT an edge in efficiency over conventional solvers like z3 and mathsat. However, this also restricts the application domains of goSAT.

Model validation

In the case of sat result, it is possible to intruct goSAT to externally validate the generated model using z3. This can be done by providing parameter -c. For example,

./gosat -c -f formula.smt2

So far, we have not encountered any unsound result. Please report to us if you find any such cases.