Skip to content

A Reinforcement-Learning-Based C Program Mutator for LLVM/GCC Bug Isolation

Notifications You must be signed in to change notification settings

haoyang9804/RecBi

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

17 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

RecBi

RecBi is a new-born technique proposed in paper Enhanced Compiler Bug Isolation via Memoized Search(written by Junjie Chen, Haoyang Ma and Lingming Zhang) for compilers' fault localization.

In this paper, we utilize two popular C compilers, LLVM and GCC, to experiment on the performance of RecBi. All our experiments are based on GCC svn repository and LLVM svn repository. Because GCC svn repository has been converted to git, we could not provide reproducible experiments about GCC bugs. In this project, we only implement the llvm part of RecBi. GCC part and LLVM part with git will be implemented some day in the future.


Here follows some steps for installing, deploying and running RecBi in your Linux workstation.

Installing

git clone https://github.com/haoyang9804/RecBi.git

In this step, do not place sudo before git. Permission problem will disable RecBi project.

Then, run chmod -R 777 RecBito circumvent permission problem.

Deploying -1: preparing for installing target LLVM trunk

In llvmbugs.txt, you can add LLVM bugs from bugzilla according to the following format:

bugId,LLVM trunk revision number,the optimization level that would not trigger the bug, the optimization level that would trigger the bug,bug type,installing situation

An example of this format is shown in llvmbugs.txt of this repository:

15920,r181189,-O2,-O3,checkIsPass_wrongcodeOneline,install_no

install_no means this LLVM trunk has not been installed while install_yes means the opposite;

bug type is defined in RecBi/llvm/failmessage_llvm_12_30.py, in which we categorized experimental bugs into 4 types:

  • checkIsPass_wrongcodeOneline: the given program in bugzilla outputs one line wrongly when the bug is triggered. Bug 15920 is of this category.
  • checkIsPass_zeroandsegmentoneline: the given program in bugzilla will crash when the bug is triggered, otherwise, it will output nothing . Bug 16041 is of this category.
  • checkIsPass_multilineswrongcode: the given program in bugzilla outputs multi-lines wrongly when the bug is triggered. Bug 25831 is of this category.
  • checkIsPass_zeroandonenumber: the given program in bugzilla will output one number when the bug is triggered, otherwise, it will output nothing. Bug 26256 is of this category

After installing the target llvm trunks, install_no will be replaced by install_yes and it will never be installed again unless you change this installing situation back.

For the information of all experimental bugs, you can check benchmark folder.

Deploying -2: configure bug information

To finish bug information configuration, you should first know whether your target bugs are in our experimental bugs list by running setup.py and it will autonomously configure the information if the answer is yes, otherwise, our code will warn you and you should configure by yourself following the step below:

  • go to ../RecBi-workplace/llvmbugs/info/bugId/* and create these two files:

    1. fail.c: the file recording the given program which can trigger the bug from bugzilla
    2. locations: the file recording 1) the LLVM trunk revision number corresponding to this bug; 2) the revision number of one LLVM trunk which fixes this bug; 3) modified files and methods(These files and methods are the reason why this bug is triggered).

    Information needed in these files could be obtained from llvm bugzilla and http://llvm.org/viewvc/llvm-project/

    An example(bug 15920) of locations is shown below

    test trunk:181189
    fixed revision:183035
    buggy locations
    file:llvm/trunk/lib/Transforms/Vectorize/LoopVectorize.cpp;method:LoopVectorizationLegality::canVectorizeInstrs()
    

Running

RecBi has two mode shown in RecBi/config/config.ini: verification and utilization.

verification is for reproducing our experiments and the output of this mode is the metrics value according to our paper.

utilization targets at isolating unknown bugs with our technique. The output of this mode is a file named rankFile.txt in ../RecBi-workplace/llvmbugs/, which records suspiciousness value of each involving compiler files in descending order. The bigger the suspiciousness value of the file, the more possible it causes the bug. You can change mode in RecBi/config/config.ini

RecBi project will run all bugs written in llvmbugs.txt unless you feed some bugIds separated by comma to reduced in RecBi/config/config.ini

Eventually, python llvm-run.py and get the output. The process of RecBi may take a while.

Welcome more people to enrich our benchmark for further-step research.

Contact

Authors information:

Name Email Adderess Github Id
Junjie Chen [email protected] JunjieChen
Haoyang Ma [email protected] haoyang9804
Lingming Zhang [email protected] lingming

About

A Reinforcement-Learning-Based C Program Mutator for LLVM/GCC Bug Isolation

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published