[klee-dev] using klee with Siemens Benchmarks
Urmas Repinski
urrimus at hotmail.com
Thu Mar 7 16:06:31 GMT 2013
Hi David.
No-no lets leave regression testing for now.
Adding klee_assume_int(char*) and klee_assume_float(char*) into klee command can really return some impressive results, will use only C arguments as strings, and will generate possibly much higher amount of test cases in this case.
What i got when i used design symbolic arguments for klee i got a lot of garbage in inputs, but there were some integer numbers also.
Later on in my design i use atoi(...) C function, get integers, and will process them as usual variables.
Klee will generate integers as strings, and possibly result will be what i need.
I just didn't found klee_assume_int(char *) and klee_assume_float(char*) in klee library, but actually i can create them by myself, and try to evaluate klee with design's symbolic string arguments.
I think this is good idea, i have to try it for now.
Thank You very much David,
Urmas Repinski
From: david.lightstone at prodigy.net
To: urrimus at hotmail.com
Subject: RE: [klee-dev] using klee with Siemens Benchmarks
Date: Thu, 7 Mar 2013 10:54:13 -0500
Urmas With respect to (2) you did not understand my intent By minimal set of regression tests I meant the following. Consider a test suite, that was used to test a program. The program has gone into distribution. Subsequently, someone asks for a change to that program. Consider the questions –(1) Which tests do not need to be re-run?(2) Which tests need to be re-run? (complement of first question)(3) Which tests will fail because of the change?(4) Which tests need to be added? The collection of tests which need to be re-run, added or corrected is the minimal regression suite (in some sense) My guess that these questions are equivalent to Error Localization. That is if there are errors, created as a consequence to the change how will they manifest themselves Dave Lightstone From: klee-dev-bounces at imperial.ac.uk [mailto:klee-dev-bounces at imperial.ac.uk] On Behalf Of Urmas Repinski
Sent: Thursday, March 07, 2013 9:15 AM
To: klee
Subject: Re: [klee-dev] using klee with Siemens Benchmarks Hi, David.(1) You appear to be investigating the effectiveness of test case generation tools. (ie fault detection effectiveness)Is this a correct conclusion?
The problem is that i were working with Error Localization and Error Corrcetion of DIAMOND FP7 project (http://fp7-diamond.eu/).
Papers will be soon ready, Model-Based Error Localization and Mutation-Based Error Correction.
But i have much time of my PhD studies left, and the superviser asked me to investigate KLEE and i think i can investigate also Error Generation possibilities using the FORENSIC tool, that were developed within contents of the project.
Tool - http://informatik.uni-bremen.de/agra/eng/forensic.php(2) You also appear to be attempting to identify a minimal set of regression test (when you have multiple instances in the history of the program)
Is comparing programs as you intend even feasible (ie time O(number of differences), O( polynomial-expression(number of differences) ) or O(exponential-expression(number of differences) )?
No, i have set of designs - siemens benchmarks, and planning to make experiments only with them, they are all very different (text processing, logical-mathematical operations, Compound C data structures, file processing) and surely complicated enough and suitable for experiments with Error Localization, Error Correction and possible Test generation algorithms.
I do not want to go deeper into code complexity, this is a little different direction i think.
(3) You also appear to be investigation fault isolation strategies (“..but i need to make inputs detect various states on the program, various errors.”)
That is a much harder problemYes, this is what i want to investigate and implements with Test Generation tool. So i have now two possibilities - i can install KLEE source on my machine and try to extend it, to make possible to generate at least something close to MC/DC test complexity, or
I can start implementing test generation from scratch on the FORENSIC tool, to re-implement known algorithms and to make experiments.
First approach seems more reasonable, SMT-based systems are widely used for Test Generation.
So at the moment i am planning to go deeper in klee and i think it possible to add paths, that differ not only by their activated components - that result high coverage, but by the entire values of the activated variables also - this will make inputs closer to MC/DC.
Please tell me what do you think about this?
Urmas Repinski. From: david.lightstone at prodigy.net
To: urrimus at hotmail.com
CC: klee-dev-bounces at imperial.ac.uk
Subject: RE: [klee-dev] using klee with Siemens Benchmarks
Date: Thu, 7 Mar 2013 08:35:15 -0500Urmas Repinski As I previously indicated KLEE is not capable of achieving MCDC coverage. That should be enough of an valuation for purposes to test effectiveness (2) You also appear to be attempting to identify a minimal set of regression test (when you have multiple instances in the history of the program)
Is comparing programs as you intend even feasible (ie time O(number of differences), O( polynomial-expression(number of differences) ) or O(exponential-expression(number of differences) )? (3) You also appear to be investigation fault isolation strategies (“..but i need to make inputs detect various states on the program, various errors.”)
That is a much harder problem Dave Lightstone
From: klee-dev-bounces at imperial.ac.uk [mailto:klee-dev-bounces at imperial.ac.uk] On Behalf Of Urmas Repinski
Sent: Thursday, March 07, 2013 7:34 AM
To: klee
Subject: Re: [klee-dev] using klee with Siemens Benchmarks Yes, Thank You.
1. How do you define 'design structure'?
This depends on tool what i am using for test generation. KLEE uses symbolic states and branches, i have also design's model - parsed design representation using hammock graph structure, any tool what generates inputs will have own design representation, and generated tests will depend on the design's representation's structure, more detailed model will generate more complete inputs.2. You might be criticising an apple for not being a good orange. KLEE is designed to tests the program it's being executed on, and if one also wants to test 41 variants, he/she will run KLEE again on the variants. I'm not sure what you would expect, even in a simple scenario like ....
Regression testing requires 2 versions of the design - correct one and faulty one, and generates inputs, based on detecting behavioral differences between two versions of the design, if i am not mistaken.
I am, from the other hand, want to use only one version of the design, and to generate as complete inputs as possible.
KLEE is pretty fast and reliable in terms of coverage, but i need to make inputs detect various states on the program, various errors.
According to http://pleuma.cc.gatech.edu/aristotle/Tools/subjects/ in Siemens Benchmarks Regression testing were used, all faulty versions of the design were processed and complete set of inputs, that detect already known errors were generated.
This works well if i have set of faulty versions, but what if not, if i have only one version of the design and want to generate complete inputs for it?
Any help will be appreciated,
Urmas RepinskiSubject: Re: [klee-dev] using klee with Siemens Benchmarks
From: paul.marinescu at imperial.ac.uk
Date: Thu, 7 Mar 2013 10:52:56 +0000
CC: klee-dev at imperial.ac.uk
To: urrimus at hotmail.com
1. How do you define 'design structure'?2. You might be criticising an apple for not being a good orange. KLEE is designed to tests the program it's being executed on, and if one also wants to test 41 variants, he/she will run KLEE again on the variants. I'm not sure what you would expect, even in a simple scenario like if (input = 100) { //something} You would need to generate 2^sizeof(input) inputs to take into account all possible mutations to the literal 100 using regression testing. Paul On 7 Mar 2013, at 05:06, Urmas Repinski <urrimus at hotmail.com> wrote: Hello.
Yes.
So KLEE takes into account only the coverage, if all paths are covered, but does not take into account entire design structure.
The errors provided by siemens benchmarks are in the operators (+ -> -, > -> <, etc) and in numbers (2 -> 3, 70 -> 80), and with klee inputs even if corresponding nodes in the design's model representation are activated, error is not propagated to the output.
Its a pity, if it is possible to test the entire structure with klee inputs (activate different bits in variables with inputs, confirm that operators values take influence on the output) then klee generated inputs will be more useful for practical imlementation.
It is possible to investigate error types, found in siemens benchmarks, and improve test generation with klee.
I am writing one article about the errors in the designs at the moment, it will be published in the september, and then i will send it to the list too.
Maybe this will make possible to improve input generation and make klee more usable in actual industry.
Urmas RepinskiSubject: Re: [klee-dev] using klee with Siemens Benchmarks
From: paul.marinescu at imperial.ac.uk
Date: Wed, 6 Mar 2013 21:19:54 +0000
CC: klee-dev at imperial.ac.uk
To: urrimus at hotmail.com
Hello Urmas,It's not that clear what you're trying to do but I assume you want to use the inputs generated by KLEE on one program version to test 41 variants. I'm afraid this might not be that easy. KLEE actually gets nearly 100% statement and branch coverage on the original program ('nearly' because there's some unreachable code). You should run it without the POSIX runtime and uclibc because you're not using them anyway. What you get is: --------------------------------------------------------------------------| Path | Instrs | Time(s) | ICov(%) | BCov(%) | ICount | Solver(%) |--------------------------------------------------------------------------| klee-out-0 | 5951 | 0.26 | 99.36 | 94.44 | 311 | 91.92 |-------------------------------------------------------------------------- From what I can tell, the path coverage is also 100%, counting just feasible paths, so there are no other inputs to generate as far as KLEE is concerned. Paul On 6 Mar 2013, at 12:31, Urmas Repinski <urrimus at hotmail.com> wrote: Hello.
My name is Urmas Repinski, PhD student in Tallinn University of Technology, Estonia.
I am trying to generate inputs for Siemens Benchmarks with KLEE.
Siemens Benchmarks - C designs with various erroneous versions for testing error localization and error correction in C designs.
They can be located and downloaded from http://pleuma.cc.gatech.edu/aristotle/Tools/subjects/
I am testing KLEE with tcas design, it is smallest and simplest.
Siemens benchmarks provide their inputs, tcas/script/runall.sh script has 1607 lines of execution of the design, and it is easy extract 1607 various inputs for the design.
But i have no idea how the inputs were generated, attaching siemens benchmarks inputs to the letter (file INPUTS_tcas_Siemens_arg).
I want to compare Siemens inputs with inputs, generated by klee.
I had installed klee as it is written in documentation - http://klee.llvm.org/GetStarted.html - with uclibc support.
Installed llvm-gcc-4.2, llvm-gcc-4.5, originally provided by Linux Mint 12 uses too new version of gcc, and this generated error then used llvm-gcc, but ok, this error solved.
When i take tcas/v1/tcas.c design, adding corresponding modification to generate inputs with klee (tcas_original.c and tcas_modified.c are attached to the letter), i get 45 inputs generated, and after modifying klee output to suitable format i get klee outputs (file KLEE_OUTPUT_arg).
urmas-PBL21 src # llvm-gcc --emit-llvm -c -g tcas.c
urmas-PBL21 src # klee --libc=uclibc --posix-runtime tcas.o
KLEE: NOTE: Using model: /home/urmas/forensic/thirdparty-install/klee/Release+Asserts/lib/libkleeRuntimePOSIX.bca
KLEE: output directory = "klee-out-143"
KLEE: WARNING: undefined reference to function: fwrite
KLEE: WARNING ONCE: calling external: syscall(16, 0, 21505, 54549392)
KLEE: WARNING ONCE: calling __user_main with extra arguments.
KLEE: done: total instructions = 14790
KLEE: done: completed paths = 45
KLEE: done: generated tests = 45
This outputs have same coverage than Siemens Inputs, but most of errors in erroneous designs are simply not detected, while Simenes inputs detect all errors.
Inputs KLEE: Coverage - 92.5532%. Total Inputs - 45 Detected Errors - 8/41
Inputs Siemens Benchmarks: Coverage - 93.0851%. Total Inputs - 1607, Detected Erroneous designs - 41/41
Maybe i have something wrong with KLEE arguments when i execute klee, can somebody help me with right klee execution?
I had tested klee with various options, but i still have 45 generated inputs. Is it possible to increase somehow number of generated inputs with klee?
Maybe WARNING KLEE: WARNING: undefined reference to function: fwrite aborts executions somewhere, but there is no fwrite function in tcas design.
Urmas Repinski<INPUTS_tcas_Siemens_arg><tcas_original.c><tcas_modified.c><KLEE_OUTPUT_arg>_______________________________________________
klee-dev mailing list
klee-dev at imperial.ac.uk
https://mailman.ic.ac.uk/mailman/listinfo/klee-dev _______________________________________________
klee-dev mailing list
klee-dev at imperial.ac.uk
https://mailman.ic.ac.uk/mailman/listinfo/klee-dev
-------------- next part --------------
HTML attachment scrubbed and removed
More information about the klee-dev
mailing list