SVMRanker usage
In the following, we assume that current directory is SVMRanker.
After having installed the required software, SVMRanker can be used by entering the src/ directory and then calling SVMRanker as follows:
python3 ./CLIMain.py --help
You should be able to see the following output.
SVMRanker --- Version 1.0
Usage: CLIMain.py [OPTIONS] COMMAND [ARGS]...
"python3 CLIMain.py COMMAND --help" for more details
Options:
--help Show this message and exit.
Commands:
lmulti
lnested
parseboogie
parsectoboogie
parsectopy
As we can see, SVMRanker provides five commands. The first two commands allow for proving termination of a given program while the remaining three can be used for parsing the input file and translate it to a different format. In the remaining part of the section we focus on the details for the use of the lmulti and lnested commands.
lmulti, short for learning multiphase ranking function, instructs SVMRanker to learn a multiphase ranking function for the given program. To get the detailed usage information for this command, one can use the following command:
python3 ./CLIMain.py lmulti --help
which produces the following output:
SVMRanker --- Version 1.0
Usage: CLIMain.py lmulti [OPTIONS] SOURCE
Options:
--depth_bound INTEGER depth bound default set to 2
--filetype [C|BOOGIE] --file C: input is c file.
--file BOOGIE:input is boogie file. default set to BOOGIE
--sample_strategy [ENLARGE|CONSTRAINT]
--sample_strategy ENLARGE: enlarge the sample zone when sample num not enough.
--sample_strategy CONSTRAINT: find feasible points by constraint if sample
num not enough default set to ENLARGE
--cutting_strategy [NEG|MINI|POS]
use f(x) < b to cut
--cutting_strategy POS: b is a postive number
--cutting_strategy NEG: b is a negative number
--cutting_strategy MINI: b is the minimum value of sampled points
default set to MINI
--template_strategy [SINGLEFULL|FULL]
templates used for learning
--template_strategy SINGLEFULL: templates are either single variable
or combination of all variables
--template_strategy FULL: template is combination of all variables
default set to SINGLEFULL
--print_level [DEBUG|INFO|NONE]
--print_level DEBUG: print all the information of the learning and debugging
--print_level INFO: print the information of the learning
--print_level NONE: only print the result information of the learning
default set to NONE
--help Show this message and exit
As the help shows, there are several options available to tune the execution of lmulti; we present their usage by means of a couple of examples.
int main() {
int x, y;
while (x > 0 || y > 0) {
x = x + y - 1;
y = y - 1;
}
}
The is the first C program we consider here; see the file src/example/Example1.c. This program can be shown to be terminating by means of a 2-multiphase ranking function, as we get by running SVMRanker to learn its multiphase ranking function as follows.
python3 ./CLIMain.py lmulti --filetype C example/Example1.c
SVMRanker completes the analysis by returning a 2-multiphase ranking function for the Example1.c program, as shown below.
SVMRanker --- Version 1.0
example/**Example1.c**
--------------------LEARNING MULTIPHASE SUMMARY-------------------
MULTIPHASE DEPTH: 2
LEARNING RESULT: TERMINATE
-----------RANKING FUNCTIONS----------
5.0 * 1 + 1.0 * y^1 + 5.0 * 1
0.0796 * x^1 + 0.482 * 1 + 0.482 * 1
Notice that we used the option –filetype to specify the type of the input program, given that SVMRanker supports both Boogie programs and C programs as input file, with the former being the default format. Furthermore, we can also provide the option –depth_bound to set the maximal number of phases SVMRanker can use when learning a multiphase ranking function. The default value of this option is 2, and this has been enough in the analysis of Example1.c, since Example1.c can be proved to terminate by a 2-multiphase ranking function. Such a default value is suitable for several programs, but it can be increased as needed, as we will see with the second example src/example/Example2.c, as shown below.
int main() {
int x, y;
while (x > 0) {
x = x + y;
y = y + z;
z = z - 1;
}
}
If we run SVMRanker on Example2.c with the default value of 2 for –depth_bound, we obtain Unknown as result. In order to get an appropriate multiphase ranking function for Example2.c, one needs to call lmulti with the option –depth_bound set to at least 3:
python3 ./CLIMain.py lmulti –filetype C –depth_bound 3 example/Example2.c
With the help of –depth_bound, SVMRanker produces the result shown below.
SVMRanker --- Version 1.0
example/Example2.c
--------------------LEARNING MULTIPHASE SUMMARY-------------------
MULTIPHASE DEPTH: 3
LEARNING RESULT: TERMINATE
-----------RANKING FUNCTIONS----------
2.0 * 1 + 2.0 * 1 + 1.0 * z^1 + 2.0 * 1
1.0 * 1 + 0.2154 * y^1 + 1.0 * 1 + 1.0 * 1
0.0911 * x^1 + 0.3226 * 1 + 0.3226 * 1 + 0.3226 * 1
We now present the other options that let SVMRanker use different strategies in the process of learning a multiphase ranking function; different strategies regarding how program data points are sampled, how the state space is cut, and what templates are used, influence the running time of SVMRanker and possibly the final result.
- –sample_strategy: This option controls the strategy SVMRanker uses to sample program data points. Possible values are CONSTRAINT and ENLARGE (the default): CONSTRAINT samples randomly the points satisfying the loop condition/guard; This strategy can be slow as we need to get the assignments of variables satisfying the loop guard. ENLARGE samples uniformly the points in a predefined space, which can be very efficient; However, if all the sampled points cannot satisfy the loop guard, we will then try to enlarge the sampling space until seeing some point satisfying the loop guard.
- –cutting_strategy: This option controls the bound b of the constraint f(x) < b that is used to cut the program state space in two parts for the current phase’s decreasing function f. Possible values are NEG, POS, and MINI (the default): NEG chooses randomly a negative value for b; POS chooses randomly a positive value for b; MINI selects the minimum value of f on the sampled points.
- –template_strategy: This option controls what templates are used in the learning procedure. Possible values are FULL and SINGLEFULL (the default): FULL uses as templates the linear combinations of all program variables; SINGLEFULL extends FULL with templates using only one variable at a time.
- –print_level: This option controls the verbosity of the SVMRanker output.
The SVMRanker command lnested, short for learning nested ranking function, is used for learning a nested ranking function for a given program. The usage information of lnested can be obtained by the following command:
python3 ./CLIMain.py lnested --help
which gives as output
SVMRanker --- Version 1.0
Usage: CLIMain.py **lnested** [OPTIONS] SOURCE
Options:
--depth_bound INTEGER depth bound default set to 2
--filetype** [C|BOOGIE] --file C: input is c file.
--file BOOGIE: input is boogie file. default set to BOOGIE
--sample_strategy [ENLARGE|CONSTRAINT]
--sample_strategy ENLARGE: enlarge the sample zone when sample num not enough.
--sample_strategy CONSTRAINT: find feasible points by constraint
if sample num not enough default set to ENLARGE
--print_level [DEBUG|INFO|NONE]
--print_level DEBUG: print all the information of the learning and debugging
--print_level INFO: print the information of the learning
--print_level NONE: only print the result information of the learning
default set to NONE
--help Show this message and exit
As we can see, the options of lnested are also the ones of lmulti_; also the use of __lnested is similar to the one of lmulti, just the outcome can be different.
For instance, we can prove termination of Example2.c by means of a learned nested ranking function by running SVMRanker as follows.
python3 ./CLIMain.py lnested --filetype C --depth_bound 3 example/Example2.c
which outputs
SVMRanker --- Version 1.0
example/Example2.c
--------------------LEARNING NESTED SUMMARY-------------------
NESTED DEPTH: 3
LEARNING RESULT: TERMINATE
-----------RANKING FUNCTIONS----------
1.0 * z^1.0 + 0.9 * 1; 1.0 * y^1.0 + 0.9 * 1; 1.0 * x^1.0 + 0.7 * 1