Windows	Ubuntu	OS X
TBD	Build	Build

About

LinearArbitrary-SeaHorn is a CHC solver for LLVM-based languages.

License

LinearArbitrary-SeaHorn is distributed under license.txt.

LinearArbitrary-SeaHorn Dependencies

• LLVM 3.6 (the build script can automatically install it)
• Z3 4.4.0 (the build script can automatically install it)
• Boost 1.63 (brew installation available on Mac os x)
• libncurses (brew installation available on Mac os x)
• libgmp (brew installation available on Mac os x)

LinearArbitrary-SeaHorn Installation

• cd seahorn ; mkdir build ; cd build
• cmake -DCMAKE_INSTALL_PREFIX=run ../
• cmake --build . to build dependencies (Z3 and LLVM)
• cmake --build . --target extra to download extra packages
• cd ../llvm-seahorn/ && git reset --hard 39aa187 && cd ../llvm-dsa/ && git reset --hard fedb3e3 && cd ../sea-dsa/ && git reset --hard 246f0f5 && cd ../crab-llvm/ && git reset --hard e2fac87 && cd ../build/ && make .. to configure extra packages
• cmake --build . --target crab && cmake .. to configure crab-llvm
• cmake --build . --target install to build LinearArbitrary-SeaHorn and install everything in run directory

LinearArbitrary-SeaHorn and dependencies are installed in build/run. LinearArbitrary-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 comes with LinearArbitrary-SeaHorn (currently 3.6). The easiest way to provide the right version of Clang is to download it from llvm.org.

Learning Libraries Installation

LinearArbitrary-SeaHorn does not explicitly link to the SVM and DT learning libraries. The interaction is currently handled by IO. This is a severe performance burden and will be improved in the future.

Installation steps are:

• cd libsvm; make clean; make
• cd C50; make clean; make

Usage

LinearArbitrary-SeaHorn provides a python script called run.py to interact with users. Given a C program annotated with assertions, users just need to type: python run.py file.c Given a directory of C programs, users just need to type: python run.py dir

To get detailed results, users can type: python run.py screen file.c

For each program, LinearArbitrary-SeaHorn outputs invariants if all assertions hold or otherwise a counterexample if any of the assertions is violated.

LinearArbitrary-SeaHorn translates file.c into LLVM bitcode, generates a set of verification conditions (VCs), and finally, solve them. The main back-end solver is Z3 + libsvm + C5.0_decision_tree.

run.py is essentially a wrapper to seahorn/build/run/bin/sea which calls the seahorn verification tool on file.c. The sea verification engine accepts several options to configure verification. Tracking how to use these options is interesting but run.py comes with a default configuration.

run.py comes with a default timeout parameter (150s) defined in the main function of the python file. Users can change it by modifying the parameter value in the source.

• PIE benchmarks

  python run.py test/c/pie

  (Our result: Among the total 117 benchmarks LinearArbitrary successfully verified 117.)

• DIG benchmarks

  python run.py test/c/pie/hola

  (Our result: Among the total 46 benchmarks LinearArbitrary successfully verified 46.)

• Recursive benchmarks

  python run.py test/c/recursions

  (Our result: Among the total 101 benchmarks LinearArbitrary successfully verified 91.)

• C Recursive+Loop benchmarks

  python run.py test/c/

  (Our result: Among the total 357 benchmarks LinearArbitrary successfully verified 345.)

• SVComp benchmarks

  python run.py test/sv-benchmarks/

  (Our result: Among the total 795 benchmarks LinearArbitrary successfully verified 765.)

• Other interesting benchmarks

  python run.py test/demo

  (Our result: Among the total 26 benchmarks LinearArbitrary successfully verified 26.)

  python run.py test/dagger

  (Our result: Among the total 25 benchmarks LinearArbitrary successfully verified 25.)

• Users are recommended to access to test_logs directory to find the details of our own experimental results.

The above arguments to run.py refer to directories. All .c files under the given directories are verified. The result of whether a verification task is successful is displayed in the command-lines window. Additionally, invariants learned to prove a program or a counterexample generated to disprove a program, for each verification task, are available in the log file ./result.log. Users can access to this file to get more insight of LinearArbitrary-SeaHorn. Alternatively one can use python run.py screen test/demo to get the tool's output directly displayed in the command-lines window.

In case users are interested in knowing what options are passed to LinearArbitrary-SeaHorn for verification in run.py, result.log is a good place for such information. We explain these options in run.py and below.

--horn-answer show learned invariants in the output

--horn-stats show some stats about learning and verification

--horn-ice use inductive counterexample guided learning to solve a verification task

--horn-ice-svm-c-paramter={int} adjust the C parameter of SVM learning

--horn-ice-c5-exec-path={str} specify where to find the decision tree learning library

--horn-ice-svm-exec-path={str} specify where to find the libsvm library

--horn-ice-mod={1,0} allow learning to use mod(%) operation of a numeric variable against a constant as a feature

--horn-ice-svm-coeff-bound={int} Z3 sometimes times-out on a formula with very large coefficients. Setting this parameter can avoid considering such candidate invariants.

--horn-rule-sample-length={int} When the verifier finds a positive sample, using this parameter allows the system to use bounded DFS on CHCs to generate more related positive samples to accelerate verification.

--horn-rule-sample-width={int} When the verifier finds a positive sample, using this parameter allows the system to use bounded BFS on CHCs to generate more related positive samples to accelerate verification.

--horn-ice-svm-call-freq-pos={int} --horn-ice-svm-call-freq-neg={int} This is an optimization implemented in the tool. The learning algorithm is based on the interaction between svm and decision tree (DT) learning. Since svm provides attributes to construct DT, if svm is called everytime a counterexample comes, DT construction may diverge since attributes change too fast. If this parameter is set to n, svm is called every n samples.

--horn-ice-local-strengthening={1,0} This is an optimization implemented in the tool. It can improve the tool's performance on some benchmarks. For a CHC, p1(x) /\ ... -> p2(x) where p1 = p2, this optimization only updates the solution of p1 during CEGAR iterations and p2 is only updated to the solution of p1 once the new p1 solution is strong enough to imply the old p2 solution.

Example

This is an example of a C program annotated with a safety property:

    /* verification command: python run.py test.c */
    #include "seahorn/seahorn.h"

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

LinearArbitrary-SeaHorn assumes all variables are initialized. Nondeterministic variables can be initialized with an unknown function. LinearArbitrary-SeaHorn follows SV-COMP convention of encoding error locations by a call to the designated error function __VERIFIER_error(). sassert() method is defined in seahorn/seahorn.h.

The solver returns invariants when __VERIFIER_error() is unreachable, and the program is considered safe. It returns counterexamples when __VERIFIER_error() is reachable and the program is unsafe. Such information can be found in result.log.

Alternatively, if running using python run.py screen test.c, the following output can be directly found in the command-line window:

************** CHCs Solved in 4.492560e-01 (secs) **************

************** Program is correct **************
REL: (verifier.error V_0 V_1 V_2) -- invariant size: 1, 2, 
REL: (main@entry V_0) -- invariant size: 1, 
REL: (main@.preheader V_0 V_1 V_2) -- invariant size: 2, 
REL: (main@.lr.ph V_0 V_1 V_2 V_3) -- invariant size: 3, 
REL: main@verifier.error.split -- invariant size: 1, 
************** Program Correctness End **************

************** Learning Statistics **************:
Total CHC size: 13
Total Relation size: 5
Total Var size: 44
Neg sample size: 11
Pos sample size: 2
Total sample size: 13
Iteration number: 15
************** Learning Statistics End **************

unsat
Function: main
main@entry: true
main@.preheader: (!(main@%k.0.i5<=0))
main@.lr.ph:
    (!(main@%i.0.i3<=0))
  (!(((1*main@%k.1.i1)+(-1*main@%j.0.i2))<=0))
main@verifier.error.split: false


************** BRUNCH STATS ***************** 
BRUNCH_STAT Result TRUE
BRUNCH_STAT Horn 0.01
BRUNCH_STAT HornClauseDB::loadZFixedPoint 0.00
BRUNCH_STAT HornifyModule 0.00
BRUNCH_STAT ICE inv 0.35
BRUNCH_STAT LargeHornifyFunction 0.00
BRUNCH_STAT seahorn_total 0.37
************** BRUNCH STATS END ***************** 
Success!
************************************************************************************

Among the total 1 benchmarks LinearArbitrary successfully verified 1.

The first line of the output gives the CHC solving time while the last line of the output only estimates the time spent within seahorn not counting learning time. If a program is correct, the tool first outputs the size of each invariant learned for a CHC unknown symbol. It then outputs some information about CHC system size and number of positive and negative samples and CEGAR iterations used to derive the solution. After that, the tool prints the learned invariants as shown above followed with some brunch stats.

Appreciation

To people who developed the SeaHorn verification framework:

• Arie Gurfinkel
• Jorge Navas
• Temesghen Kahsai

To people who developed an initial implementation of the ICE algorithm inside SeaHorn:

• Chenguang Zhu