seahorn

About

SeaHorn is an automated analysis framework for LLVM-based languages. This version compiles against LLVM 14.

Some of the supported features are

• Abstract Interpretation-based static analysis
• Unification-based Context-Sensitive pointer analysis
• SMT-based Bounded Model Checking (i.e., symbolic execution)
• CHC-based Software Model Checking (i.e., invariant inference)
• Executable counterexamples (i.e., no reports, just bugs!)

SeaHorn is developed primarily as a framework for conducting research in automated verification. The frameworks provides many components that can be put together in a variety of ways. However, it is not an "out-of-the-box" static analysis tool.

Many analysis tools and examples are provided with the framework. We are constantly looking for new applications and provide support to new users. For more information on what is happening, check our (infrequently updated) blog.

License

SeaHorn is distributed under a modified BSD license. See license.txt for details.

Introduction

SeaHorn provides a python script called sea to interact with users. Given a C program annotated with assertions, users just need to type: sea pf file.c

The result of sea-pf is unsat if all assertions hold, an sat if any of the assertions are violated.

The option pf tells SeaHorn to translate file.c into LLVM bitcode, generate a set of verification conditions (VCs), and finally, solve them. The main back-end solver is spacer.

The command pf provides, among others, the following options:

• --show-invars: display computed invariants if answer was unsat.

• --cex=FILE : stores a counter-example in FILE if answer was sat.

• -g : compiles with debug information for more trackable counterexamples.

• --step=large: large-step encoding. Each transition relation corresponds to a loop-free fragments.

• --step=small: small-step encoding. Each transition relation corresponds to a basic block.

• --track=reg : model (integer) registers only.

• --track=ptr : model registers and pointers (but not memory content)

• --track=mem: model both scalars, pointers, and memory contents

• --inline : inlines the program before verification

• --crab : inject invariants in spacer generated by the Crab abstract-interpretation-based tool. Read here for details about all Crab options (prefix --crab). You can see which invariants are inferred by Crab by typing option --log=crab.

• --bmc: use BMC engine.

sea pf is a pipeline that runs multiple commands. Individual parts of the pipeline can be run separately as well:

1. sea fe file.c -o file.bc: SeaHorn frontend translates a C program into optimized LLVM bitcode including mixed-semantics transformation.

2. sea horn file.bc -o file.smt2: SeaHorn generates the verification conditions from file.bc and outputs them into SMT-LIB v2 format. Users can choose between different encoding styles with several levels of precision by adding:

   • --step={small,large,fsmall,flarge} where small is small step encoding, large is block-large encoding, fsmall is small step encoding producing flat Horn clauses (i.e., it generates a transition system with only one predicate), and flarge: block-large encoding producing flat Horn clauses.

   • --track={reg,ptr,mem} where reg only models integer scalars, ptr models reg and pointer addresses, and mem models ptr and memory contents.

3. sea smt file.c -o file.smt2: Generates CHC in SMT-LIB2 format. Is an alias for sea fe followed by sea horn. The command sea pf is an alias for sea smt --prove.

4. sea clp file.c -o file.clp: Generates CHC in CLP format.

5. sea lfe file.c -o file.ll : runs the legacy front-end

To see all the commands, type sea --help. To see options for each individual command CMD (e.g, horn), type sea CMD --help (e.g., sea horn --help).

Static Analysis with Abstract Interpretation

Inference of Inductive Invariants using Crab

SeaHorn does not use Crab by default. To enable Crab, add the option --crab to the sea command.

The abstract interpreter is by default intra-procedural and it uses the Zones domain as the numerical abstract domain. These default options should be enough for normal users. For developers, if you want to use other abstract domains you need to:

1. Compile with cmake options -DCRAB_USE_LDD=ON -DCRAB_USE_ELINA=ON
2. Run sea with option --crab-dom=DOM where DOM can be:
   • int for intervals
   • term-int for intervals with uninterpreted functions
   • boxes: for disjunctive intervals
   • oct for octagons
   • pk for polyhedra

To use the crab inter-procedural analysis you need to run sea with option --crab-inter

By default, the abstract interpreter only reasons about scalar variables (i.e., LLVM registers). Run sea with the options --crab-track=mem --crab-singleton-aliases=true to reason about memory contents.

How to use Invariants generated by Crab in Spacer

Crab is mostly path-insensitive while Spacer, our Horn clause solver, is path-sensitive. Although path-insensitive analyses are more efficient, path-sensitivity is typically required to prove the property of interest. This motivates our decision of running first Crab (if option --crab) and then pass the generated invariants to Spacer. There are currently two ways for Spacer to use the invariants generated by Crab. The sea option --horn-use-invs=VAL tells spacer how to use those invariants:

• If VAL is equal to bg then invariants are only used to help spacer in proving a lemma is inductive.
• If VAL is equal to always then the behavior is similar to bg but in addition invariants are also used to help spacer to block a counterexample.

The default value is bg. Of course, if Crab can prove the program is safe then Spacer does not incur in any extra cost.

Property Specification

Properties are assumed to be assertions. SeaHorn provides a static assertion command sassert, as illustrated in the following example

/* verification command: sea pf --horn-stats test.c */
#include "seahorn/seahorn.h"
extern int nd();

int main(void) {
    int k = 1;
    int i = 1;
    int j = 0;
    int n = nd();
    while (i < n) {
        j = 0;
        while (j < i) {
            k += (i - j);
            j++;
        }
        i++;
    }
    sassert(k >= n);
}

Internally, SeaHorn follows SV-COMP convention of encoding error locations by a call to the designated error function __VERIFIER_error(). SeaHorn returns unsat when __VERIFIER_error() is unreachable, and the program is considered safe. SeaHorn returns sat when __VERIFIER_error() is reachable and the program is unsafe. sassert() method is defined in seahorn/seahorn.h.

Inspect Code

Apart from proving properties or producing counterexamples, it is sometimes useful to inspect the code under analysis to get an idea of its complexity. For this, SeaHorn provides a command sea inspect. For instance, given a C program ex.c type:

sea inspect ex.c --sea-dsa=cs+t --mem-dot 

The option --sea-dsa=cs+t enables the new context-, type-sensitive sea-dsa analysis described in FMCAD19. This command generates a FUN.mem.dot file for each function FUN in the input program. To visualize the graph of the main function, use web graphivz interface, or the following commands:

$ dot -Tpdf main.mem.dot -o main.mem.pdf

More details on the memory graphs is in the SeaDsa repository: here.

Use sea inspect --help to see all options. Currently, the available options are:

• sea inspect --profiler prints the number of functions, basic blocks, loops, etc.
• sea inspect --mem-callgraph-dot prints to dot format the call graph constructed by SeaDsa.
• sea inspect --mem-callgraph-stats prints to standard output some statstics about the call graph construction done by SeaDsa.
• sea inspect --mem-smc-stats prints the number of memory accesses that can be proven safe by SeaDsa.

Installation

The easiest way to get started with SeaHorn is via a docker distribution.

$ docker pull seahorn/seahorn-llvm10:nightly
$ docker run --rm -it seahorn/seahorn-llvm10:nightly

Start with exploring what the sea command can do:

$ sea --help
$ sea pf --help

The nightly tag is automatically refreshed daily and contains the latest development version. We maintain all other tags (that require manual update) infrequently. Check the dates on DockerHub and log an issue on GitHub if they are too stale.

Additional examples and configuration options are on the blog. The blog is updated infrequently. In particular, options change, features are phased out, new things are added. If you find problems in the blog, let us know. We at least will update the blog post to indicate that it is not expected to work with the latest version of the code.

You can also manually install by:

Following the instructions in the Docker file Dockerfile: docker/seahorn-builder.Dockerfile.

If this does not work, run:

$ wget https://apt.llvm.org/llvm.sh
$ chmod +x llvm.sh
$ sudo ./llvm.sh 14
$ apt download libpolly-14-dev && sudo dpkg --force-all -i libpolly-14-dev*

The first 3 commands will install LLVM 14, the 4th will install libpolly which is wrongly omitted from LLVM 14 (but included in subsequent versions)

Next, follow the instruction in the Docker file above

Developer's Zone

The information from this point on is for developers only. If you would like to contribute to SeaHorn, build your own tools based on it, or just interested in how it works inside, keep reading.

Compilation Instructions

SeaHorn requires LLVM, Z3, and boost. The exact versions of the libraries keep changing, but cmake craft is used to check that right version is available.

To specify a specific version of any of the dependencies, use the usual <PackageName>_ROOT and/or <PackageName>_DIR (see find_package() for details) cmake variables.

SeaHorn is broken into multiple components that live in different repositories (under SeaHorn organization). The build process automatically checks out everything as necessary. For current build instructions, check the CI scripts.

These are the generic steps. Do NOT use them. Read on for a better way:

1. cd seahorn ; mkdir build ; cd build (The build directory can also be outside the source directory.)
2. cmake -DCMAKE_INSTALL_PREFIX=run ../ (Install is required!)
3. cmake --build . --target extra && cmake .. (clones components that live elsewhere)
4. cmake --build . --target crab && cmake .. (clones crab library)
5. cmake --build . --target install (build and install everything under run)
6. cmake --build . --target test-all (run tests)

Note: install target is required for tests to work!

Better Compilation Instructions

For an enhanced development experience:

1. Use clang
2. On Linux, use lld linker
3. Include debug symbols in Release builds
4. Use Ninja
5. Export compile_commands.json

On Linux, we suggest the following cmake configuration:

$ cd build
$ cmake -DCMAKE_INSTALL_PREFIX=run 
      -DCMAKE_BUILD_TYPE=RelWithDebInfo 
      -DCMAKE_CXX_COMPILER="clang++-14" 
      -DCMAKE_C_COMPILER="clang-14" 
      -DSEA_ENABLE_LLD=ON  
      -DCMAKE_EXPORT_COMPILE_COMMANDS=1 
      ../ 
      -DZ3_ROOT=<Z3_ROOT> 
      -DLLVM_DIR=<LLMV_CMAKE_DIR> 
      -GNinja
$ (cd .. && ln -sf build/compile_commands.json .)

where <Z3_ROOT is a directory containing Z3 binary distribution, and LLMV_CMAKE_DIR is directory containing LLVMConfig.cmake.

Other legal options for CMAKE_BUILD_TYPE are Debug and Coverage. Note that the CMAKE_BUILD_TYPE must be compatible with the one used to compile LLVM. In particular, you will need a Debug build of LLVM to compile SeaHorn in `Debug** mode. Make sure you have plenty of patience, disk space, and time if you decide to go this route.

Alternatively, the project can be configured using cmake presets. To do this, simply run the following command:

$ cmake --preset <BUILD_TYPE>-<PRESET_NAME>

to configure cmake, where <BUILD_TYPE> is one of: Debug, RelWithDebInfo or Coverage and <PRESET_NAME> is the preset you would like to use. The presets that are currently available are: jammy. These presets assume that you have Z3 installed in /opt/z3-4.8.9 and Yices installed in /opt/yices-2.6.1.

This will also allow the project to be configured and compiled within VS Code using the CMake Tools extension.

If you would like to use different compilation settings or if you have Z3 or Yices installed in any other directory, you will need to make your own CMakeUserPresets.json file with your own presets.

Compiling on a Mac

Do not include -DSEA_ENABLE_LLD=ON. The default compiler is clang, so you might not need to set it explicitly.

The EXTRA Target

SeaHorn provides several components that are automatically cloned and installed via the extra target. These components can be used by other projects outside of SeaHorn.

• sea-dsa: git clone https://github.com/seahorn/sea-dsa.git

  sea-dsa is a new DSA-based heap analysis. Unlike llvm-dsa, sea-dsa is context-sensitive and therefore, a finer-grained partition of the heap can be generated in presence of function calls.

• clam: git clone https://github.com/seahorn/crab-llvm.git

  clam provides inductive invariants using abstract interpretation techniques to the rest of SeaHorn's backends.

• llvm-seahorn: git clone https://github.com/seahorn/llvm-seahorn.git

  llvm-seahorn provides tailored-to-verification versions of InstCombine and IndVarSimplify LLVM passes as well as a LLVM pass to convert undefined values into nondeterministic calls, among other things.

SeaHorn doesn't come with its own version of Clang and expects to find it either in the build directory (run/bin) or in PATH. Make sure that the version of Clang matches the version of LLVM that was used to compile SeaHorn (currently LLVM14). The easiest way to provide the right version of Clang is to download it from llvm.org, unpact it somewhere and create a symbolic link to clang and clang++ in run/bin.

$ cd seahorn/build/run/bin
$ ln -s <CLANG_ROOT>/bin/clang clang
$ ln -s <CLANG_ROOT>/bin/clang++ clang++

where <CLANG_ROOT> is the location at which Clang was unpacked.

Tests

Testing infrastructure depends on several Python packages. These have their own dependencies. If you cannot figure them out, use docker instead.

$ pip install lit OutputCheck networkx pygraphviz

Coverage

We can use gcov and lcov to generate test coverage information for SeaHorn. To build with coverage enabled, we need to run build under a different directory and set CMAKE_BUILD_TYPE to Coverage during cmake configuration.

Example steps for generating coverage report for the test-opsem target:

1. mkdir coverage; cd coverage create and enter coverage build directory
2. cmake -DCMAKE_BUILD_TYPE=Coverage <other flags as you wish> ../
3. Complete the build as usual
4. cmake --build . --target test-opsem Run OpSem tests, now .gcda and .gcno files should be created in the corresponding CMakeFiles directories
5. lcov -c --directory lib/seahorn/CMakeFiles/seahorn.LIB.dir/ -o coverage.info collect coverage data from desired module, if clang is used as the compiler instead of gcc, create a bash script llvm-gcov.sh:

#!/bin/bash
exec llvm-cov gcov "$@"

$ chmod +x llvm-gcov.sh

then append --gcov-tool <path_to_wrapper_script>/llvm-gcov.sh to the lcov -c ... command. 6. extract data from desired directories and generate html report:

lcov --extract coverage.info "*/lib/seahorn/*" -o lib.info
lcov --extract coverage.info "*/include/seahorn/*" -o header.info
cat header.info lib.info > all.info
genhtml all.info --output-directory coverage_report

then open coverage_report/index.html in browser to view the coverage report

Also see scripts/coverage for scripts used by the CI. Coverage report for nightly builds is available at codecov

Code indexing

Compilation database for the seahorn project and all its sub-projects is generated using -DCMAKE_EXPORT_COMPILE_COMMANDS=ON option for cmake.

An easy way to get code indexing to work with with compilation database support is to link the compilation_database.json file into the main project directory and follow instructions specific to your editor.

• Clion -- https://www.jetbrains.com/help/clion/compilation-database.html
• Vim (YouCompleteMe) -- instructions
• Emacs -- use lsp-ui with clangd which are available in spacemacs develop branch

Remote Configuration for CLion

For a detailed guide for a remote workflow with CLion check Clion-configuration.

Remote Configuration for Emacs (and other Editors)

Use our fork of mainframer. Don't miss the example configuration.

Contributors

• Arie Gurfinkel
• Jorge Navas
• Temesghen Kahsai
• Jakub Kuderski
• Nham Le
• Charles Lei
• Xiang Zhou
• Siddharth Priya
• Isabel Garcia-Contreras
• Yusen Su
• EMAIL ME IF YOUR NAME SHOULD BE HERE