LSRepair: Live Search of Fix Ingredients for Automated Program Repair
Download as PPTX, PDF1 like471 views
The document presents LSRepair, a tool for automated program repair that searches for fix ingredients in real-world code bases to address bugs. It highlights the process of fault localization and patch generation using signature, syntactic, and semantic similarity among methods. LSRepair is quantitatively compared to state-of-the-art automated program repair tools, demonstrating its effectiveness and limitations in fixing bugs correctly.
![1
> Basic Process of Automated Program Repair (APR)
Fault
Localization
Test
Pass
Fail
Patch
Candidate
APR
Tools
Suspicious
buggy code
Where is the code to be fixed? How to generate patches? Is the patch correct?
passing
tests
Passing
tests
Failing
tests
State-of-the-art APR tools for Java bugs: jGenProg [9], jKali [9], jMutRepair [9], HDRepair [10], Nopol [28], ACS [29],
ssFix [54], ELIXIR [55], SketchFix [56], CapGen [13], and SimFix1.
1. Jiang, Jiajun, et al. "Shaping Program Repair Space with Existing Patches and Similar Code." ISSTA 2018.
Plausible
Patch](https://image.slidesharecdn.com/apsec-2018-190406074322/85/LSRepair-Live-Search-of-Fix-Ingredients-for-Automated-Program-Repair-2-320.jpg)
![10
> Search Strategies for Similar Methods
……..
...
Suspicious method Search similar methods
S1. Signature-similar methods.
S2. Syntactically similar methods.
S3. Semantically similar methods.
Method Signature (MS):
MS = (rt, mn, Args),
rt: return type.
mn: method name.
Args: a set of parameter types.
Method code search syntactic-similar
methods with deep learning technique in
our previous study [52] (TSE’18).
Method code search semantic-similar
methods with FaCoY [50] (ICSE’18).](https://image.slidesharecdn.com/apsec-2018-190406074322/85/LSRepair-Live-Search-of-Fix-Ingredients-for-Automated-Program-Repair-11-320.jpg)
![12
> Patch Generation
1. Signature-similar methods.
Replacing the buggy method with signature-similar methods directly.
2. Syntactically similar methods.
3. Semantically similar methods.
Pivot statements (buggy code and patch candidate):
IfStatement, ExpressionStatement, VariableDeclaration, and ReturnStatement [53].
if (dataset != null) {
……
}
if (element == null) {
……
}
IfStatement: Modify its conditional expression by checking their differences on non-variable code.
Others: 1) Insert an IfStatement.
2) Delete the IfStatement.
3) Modify the IfStatement.
Buggy Code Code in similar methods
if (dataset == null) {
……
}
Patch Code
Type var = var2.get();
Buggy code
if (var_ != null) {
Type var = var_.get();
}
Code in similar methods
Type var = var2.get();
Buggy code
if (var_ != null) {
Type var = var_.get();
}
Code in similar methods
if (var2 == null) {
Type var = var2.get();
}
Buggy code
if (var_ == null) {
Type var = var_.get();
}
Code in similar methods](https://image.slidesharecdn.com/apsec-2018-190406074322/85/LSRepair-Live-Search-of-Fix-Ingredients-for-Automated-Program-Repair-13-320.jpg)
![15
> RQ: to what extent real-world bugs can be fixed by
LSRepair?
Project # Bugs kLoC # Tests
Chart 26 96 2,205
Closure 133 90 7,927
Lang 65 22 2,245
Math 106 85 3,602
Mockito 38 11 1,457
Time 27 28 4,130
Total 395 332 22,954
Defects4J [57] Dataset Information.](https://image.slidesharecdn.com/apsec-2018-190406074322/85/LSRepair-Live-Search-of-Fix-Ingredients-for-Automated-Program-Repair-16-320.jpg)
![17
> Quantitative Comparison with State-of-the-art APR Tools
APR Tool # plausibly fixed bugs # correctly fixed bugs
LSRepair 38 19
jGenProg [9] 29 5
HDRepair [10] 16 13
Nopol [28] 35 5
ACS [29] 21 18
ssFix [54] 60 20
ELIXIR [55] 41 26
SketchFix [56] 26 19
CapGen [13] 25 21
Chart-4,
Lang-29,46,48, 52, 54.
Math-91, 94.
Mockito-13.
Bugs are firstly and correctly
fixed by LSRepair.
LSRepair can be complementary to the state-of-the-art
automated program repair techniques.](https://image.slidesharecdn.com/apsec-2018-190406074322/85/LSRepair-Live-Search-of-Fix-Ingredients-for-Automated-Program-Repair-18-320.jpg)
LSRepair: Live Search of Fix Ingredients for Automated Program Repair
- 1. LSRepair: Live Search of Fix Ingredients for Automated Program Repair Kui Liu, Anil Koyuncu, Kisub Kim, Dongsun Kim, and Tegawendé F. Bissyandé SnT, University of Luxembourg, Luxembourg @ Nara Japan, 25th APSEC 2018December 7, 2018
- 2. 1 > Basic Process of Automated Program Repair (APR) Fault Localization Test Pass Fail Patch Candidate APR Tools Suspicious buggy code Where is the code to be fixed? How to generate patches? Is the patch correct? passing tests Passing tests Failing tests State-of-the-art APR tools for Java bugs: jGenProg [9], jKali [9], jMutRepair [9], HDRepair [10], Nopol [28], ACS [29], ssFix [54], ELIXIR [55], SketchFix [56], CapGen [13], and SimFix1. 1. Jiang, Jiajun, et al. "Shaping Program Repair Space with Existing Patches and Similar Code." ISSTA 2018. Plausible Patch
- 3. 2 > Statement Level Bug Fixing Class name Line No. Suspiciousness …..….. Suspicious statements of bug Chart-1 in Defects4J reported by GZoltar 0.1.1. …..….. Bug Chart_1 in Defects4J fixed by jMutRepair, ELIXIR, ssFix, JAID, SketchFix, CapGen, SimFix. org/jfree/chart/renderer/category/AbstractCategoryItemRenderer.java
- 4. 3 > Search Space Issues P P P P P P P P P P PP P P P P P PP P P P P P P P P P P P PP P P P P P Incorrect patch. P Correct patch. P Plausible but incorrect patch. P P (1) The search space could be too small to include correct patches. (2) The search space could be too big to include more plausible but incorrect patches. P P Search space: a set of patches generated by APR tools.
- 5. 4 > Suspicious but non-buggy statements in one method
- 6. 5 > Question Could we fix bugs at method declaration level to reduce the search space of fix ingredients for incorrect patches and purify the search space for correct patches?
- 7. 6 > Intuition Developers write code or clone code to address similar tasks. program1 program2 program3 DesignsDeveloping Routine Our intuition is that while some code may be buggy, the similar code may have been fixed.
- 8. 7 > LSRepair Approach Live Search of fix ingredients in real-world code bases to repair bugs automatically. Overview of LSRepair Suspicious methods Fault localization Search similar methods Generate and Validate patches
- 10. 9 > Fault Localization passing tests Passing tests Failing tests Gzoltar 0.1.1 Suspicious statement list Suspicious method list … ... … ... … ... …..….. Bug Chart_1 in Defects4J. org/jfree/chart/renderer/category/AbstractCategoryItemRenderer.java
- 11. 10 > Search Strategies for Similar Methods …….. ... Suspicious method Search similar methods S1. Signature-similar methods. S2. Syntactically similar methods. S3. Semantically similar methods. Method Signature (MS): MS = (rt, mn, Args), rt: return type. mn: method name. Args: a set of parameter types. Method code search syntactic-similar methods with deep learning technique in our previous study [52] (TSE’18). Method code search semantic-similar methods with FaCoY [50] (ICSE’18).
- 12. 11 > Patch Generation 1. Signature-similar methods. Replacing the buggy method with signature-similar methods directly.
- 13. 12 > Patch Generation 1. Signature-similar methods. Replacing the buggy method with signature-similar methods directly. 2. Syntactically similar methods. 3. Semantically similar methods. Pivot statements (buggy code and patch candidate): IfStatement, ExpressionStatement, VariableDeclaration, and ReturnStatement [53]. if (dataset != null) { …… } if (element == null) { …… } IfStatement: Modify its conditional expression by checking their differences on non-variable code. Others: 1) Insert an IfStatement. 2) Delete the IfStatement. 3) Modify the IfStatement. Buggy Code Code in similar methods if (dataset == null) { …… } Patch Code Type var = var2.get(); Buggy code if (var_ != null) { Type var = var_.get(); } Code in similar methods Type var = var2.get(); Buggy code if (var_ != null) { Type var = var_.get(); } Code in similar methods if (var2 == null) { Type var = var2.get(); } Buggy code if (var_ == null) { Type var = var_.get(); } Code in similar methods
- 14. 13 ASSESSMENT
- 15. 14 > Real World Code Bases 10,449 GitHub Java projects used in FaCoy (11,043,044 methods). Search space of similar methods: • Signature-similar methods: all similar methods in our data set. • Syntactically similar methods: top-10 most similar methods. • Semantically similar methods: top-10 most similar methods.
- 16. 15 > RQ: to what extent real-world bugs can be fixed by LSRepair? Project # Bugs kLoC # Tests Chart 26 96 2,205 Closure 133 90 7,927 Lang 65 22 2,245 Math 106 85 3,602 Mockito 38 11 1,457 Time 27 28 4,130 Total 395 332 22,954 Defects4J [57] Dataset Information.
- 17. 16 > Bugs correctly fixed by LSRepair Bugs in Defects4J are correctly fixed by LSRepair. S1. Signature-similar methods. S2. Syntactically similar methods. S3. Semantically similar methods.
- 18. 17 > Quantitative Comparison with State-of-the-art APR Tools APR Tool # plausibly fixed bugs # correctly fixed bugs LSRepair 38 19 jGenProg [9] 29 5 HDRepair [10] 16 13 Nopol [28] 35 5 ACS [29] 21 18 ssFix [54] 60 20 ELIXIR [55] 41 26 SketchFix [56] 26 19 CapGen [13] 25 21 Chart-4, Lang-29,46,48, 52, 54. Math-91, 94. Mockito-13. Bugs are firstly and correctly fixed by LSRepair. LSRepair can be complementary to the state-of-the-art automated program repair techniques.
- 19. 18 > Limitation and Discussion 3. Method level granularity. It cannot fix the bugs located in TypeDeclaration or FieldDeclaration. 2. Code Transformation. Our prototype uses straightforward heuristics to transform the ingredients. 1. Syntactically or Semantically Similar Code Search Problem. We use naive heuristics to prioritize code, based on the cosine similarity (e.g., syntactic and semantic) with a given buggy code.
- 20. 19 > Summary 13 > Bugs correctly fixed by LSRepair Bugs in Defects4J are correctly fixed by LSRepair. 10 > Search Strategies …….. ... Suspicious method Search similar methods 1. Signature-similar methods. 2. Syntactically similar methods. 3. Semantically similar methods. 5 > Motivation Developers write code to address similar tasks, or clone code. program1 program2 program3 DesignsDeveloping Routine Our intuition is that while some code may be buggy, the similar code may have been fixed . 6 > LSRepair Approach Live Search of fix ingredients in real-world code bases to repair bugs automatically. Fault Localization to identify faulty methods Generate patches By code transform Validate and check correctness Selection of code search strategy Program https://github.com/AutoProRepair/LSRepair
Editor's Notes
- #3: In order to reduce the manual debugging effort, various automated program repair tools have been proposed. Now, let’s take a look at the basic process of Automated program repair. Given a buggy program with its passing and failing tests, fault localization techniques are used to locate the suspicious buggy code positions. Automated program tools are designed to mutate the suspicious code to generate patch candidates. Only the patch candidate that can pass all tests is considered as a plausible patch of the buggy program. For Java program bugs, the state-of-the-art APR tools are designed in this way. This kind of bug fixing focuses on the statement level.
- #4: For example, given the bug Chart-1 in Defects4J, the fault localization tool, Gzoltar, reports the suspicious code in this way. The column means the suspicious buggy class names, the second line is the suspicious code line number in the suspicious buggy file. The last column denotes the suspiciousness value of the current code line, which is calculated with the tests running traces. Therefore, APR tools are going to mutate these suspicious statements one by one, until a plausible patch is generated, or time out. The exact the bug position of Chart-1 is Line 1797 in file …. Thus, APR tools need to try all the suspicious code positions before this one.
- #5: In this way to fix bugs, there are two critical issues which must be considered. We assume that, there is a patch universe which contains many incorrect patches, some plausible but incorrect patches, and a few correct patches. If APR tools limit the search space size, it might be too small to include correct patches. If APR tools enlarge the search space, it could enlarge the possibility of including correct patches, but it also enlarge the possibility of including more plausible patches. How to resolve or relieve these two issues is important to search based automated program repair.
- #6: Considering several statements in one method could be reported as suspicious statements, thus, could we ….
- #7: Considering several statements in one method could be reported as suspicious statements, thus, could we ….
- #8: As we known, many programs have the same or similar developing routine, data structure and designs. Thus, developers always write code and clone code to address similar tasks. Our intuition is that ….
- #9: With our intuition, we propose to repair bugs with live search of fix ingredients in real-world code bases. The overflow of LSRepair is, given a buggy program ……
- #10: Now, let’s take a look at the LSRepair design.