Automatically Finding Patches Using Genetic Programming

Transcription

1 Automatically Finding Patches Using Genetic Programming Authors: Westley Weimer, ThanhVu Nguyen, Claire Le Goues, Stephanie Forrest Presented by: David DeBonis, Qi Lu, Shuang Yang Department of Computer Science University of New Mexico February, 0

2 Outline Mutation Crossover 5

3 Error Correction in Source Code Based on positive / negative test cases Program isolation Repair by insert / delete / swap Repeat until a variant passes all tests Minimize difference

4 Example of Errant Code Listing : Euclid s greatest common divisor algorithm /* requires : a >= 0, b >= 0 */ void gcd ( int a, int b ) { if ( a == 0) { printf ( " % d ", b ); } while ( b!= 0) a = a - b; else b = b - a; printf ( " % d ", a ); exit (0); }

5 Variants Creation and Representation Restricted to code substitutions from other parts Mutation / Crossover constrained to area relevant to error Abstract Syntax Tree (AST) Weighted program path P Fitness = N i= wi passed test cases

6 An abstract syntax tree(ast). It includes all of the statements in the program. A weighted path through the program, (statement, weight). The weighted path is a list of pairs, each pair contains a statement and a weight based on that statement s occurrences in various testcases.

7 The algorithm of mutation

8 The example for Mutation 5 6 while ( b!= 0) a = a - b; else b = b + a; return a ;

9 The example for Mutation statement sequence return 0.0 while op:!= : a branch 5 6 while ( b!= 0) a = a - b; else b = b + a; return a ; : b constant: 0 : a op: > : b : a : a 0 bin op: : b : b Figure: The AST of the program : b bin op: + : a

10 Mutation: Swap statement sequence return 0.0 while op:!= : a branch : b constant: 0 : a op: > : b : b : b 0 bin op: + : a : a : a bin op: : b

11 Mutation: Swap 5 6 while ( b!= 0) a = a - b; else b = b + a; return a ; 5 6 while ( b!= 0) b = b + a; else a = a - b; return a ;

12 Mutation: Insertion statement sequence 0.0 condition while 0.0 body op:!= : a branch : b constant: 0 : a 0.0 return condition if body op: > : b : a : a bin op: else body : a 0 bin op: : b : a : b : b : b bin op: + : a

13 Mutation: Insertion 5 6 while ( b!= 0) a = a - b; else b = b + a; return a ; while ( b!= 0) a = a - b; else b = b + a; a = a - b; return a ;

14 Mutation: Deletion statement sequence 0.0 condition while return body op:!= : a branch : b constant: 0 : a condition if body op: > : b : a : a else body 0 bin op: : b : b : b Deletion bin op: + : a

15 Mutation: Deletion 5 6 while ( b!= 0) a = a - b; else b = b + a; return a ; while ( b!= 0) a = a - b; return a ;

16 Crossover Choose a cutoff point for each program Combine the first part of one program with the second part of another and vice versa Input: [P, P, P, P ] and [Q, Q, Q, Q ] with cutoff Child: [P, P, Q, Q ] and [Q, Q, P, P ]

17 Crossover a = a - b; else b = b - a; cutoff point at a = a + b; else b = b + a;

18 Crossover branch branch if-body condition else-body op: > bin op op: if-body condition op: > bin op op: else-body bin op op: + bin op op: +

19 Crossover branch branch if-body condition else-body op: > bin op op: if-body condition op: > bin op op: + else-body bin op op: + bin op op:

20 Crossover a = a - b; else b = b + a; a = a + b; else b = b - a;

21 Crossover

22 Crossover branch branch if-body condition else-body op: > bin op op: + if-body condition op: > bin op op: else-body bin op op: bin op op: +

23 Demonstrated GP approach Based on AST representation Mutate / Crossover algorithm Minimal code changes Experiments over 0 test cases Sizeable code bases Orthogonal errors present Reasonable solution found ( 50%) Efficacy to code maintenance Experiments using Amazon Cloud Equated to under $8 per bug on average Who wants to manually debug anyway!

24 Limitations and Assumptions Only as good as your test cases Not necessarily memory / computing time optimized Expects redundant code segments How well does it scale? by number of bugs... by number of revisions... by number of LOC... Will this technique work on another GP?