Deckard: Scalable and Accurate Tree-based Detection of Code Clones. Lingxiao Jiang, Ghassan Misherghi, Zhendong Su, Stephane Glondu

Transcription

1 Deckard: Scalable and Accurate Tree-based Detection of Code Clones Lingxiao Jiang, Ghassan Misherghi, Zhendong Su, Stephane Glondu

2 The Problem Find similar code in large code bases, often referred to as Clone Detection. Many applications in software engineering Refactoring Evolution study Quality assessment Bug detection y[i] = ; y[i] = ; Code 1 Code 2 2

3 Clone Detection Techniques String-based E.g., Baker (1995, 1997) Token-based E.g., CCFinder (2002), CP-Miner (2004) Tree-based E.g., CloneDR (1998, 2004) Semantic-based E.g., Komondoor & Horwitz (2001) 3

4 Limitations of Existing techniques String-based, token-based Sensitive to code format E.g., x = a*b+c vs. z = c+a*b Tree-based, semantic-based Not scalable 4

5 Deckard s Goal More robust against code modification Scalable to million-line programs 5

6 Tree Similarity and Code Clones Well defined tree similarity problem Tree editing distance δ(t 1, T 2 ): root Minimal # of edit operations fun id argu = + (insert/delete a node or change labels) e f x y a b T 1 and T 2 are σ-similar if δ(t 1, T 2 ) < σ Code clones based on δ clone pair, if δ(t 1, T 2 ) < σ y[i] = ; y[i] = ; Code 1 Code 2 6

7 Not Scalable A clone detection directly based on the definition would not scale: Computing tree editing distance is expensive = root + x y a b Minimal # of edit operations O ( T 1 * T 2 * d 1 * d 2 ) (Zhang et al. 1989) id fun e argu f where d i is the minimum of the depth of T i and the number of leaves of T i Need to compare many pairs of subtrees Program.. Code 1 Code i Code n Quadratic # of Tree Comparison 7

8 The Idea Numerical Vectors Numerical vectors are much easier to compare than trees Suppose v 1 = <x 1,,x n > and v 2 = <y 1,, y n > Hamming distance (or the l 1 norm): Euclidean distance (or the l 2 norm): Program Characterize parse trees as numerical vectors Cluster vectors based on numerical distances Code corresponding to vectors within a same cluster are clones decl var var x root = y a + id call b e argu f v 1 v i v n v 1, v k, v j, v i v n, v m,, v y C 1,C k, C j, C i,c j C n,c m, C y 8

9 Deckard s High Level Architecture 9

10 Outline Define characteristic vectors for trees Relate H, D with δ Generate vectors Evaluate Deckard Implement Deckard Cluster vectors with Locality-Sensitive Hashing 10

11 Characteristic Vectors for Trees Definitions q-level atomic patterns Labeled complete binary trees of height q E.g., 2-level + id id q 2 1 At most L patterns, where L is the set of labels q-level vector for a tree or tree-forest T 2 1 <b 1,, b k >, where k L Each b i counts # of i-th pattern in T q 11

12 An Sample Characteristic Vector 1-level patterns: <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> 12

13 An Sample Characteristic Vector 1-level patterns: <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> 13

14 Relate H(T 1,T 2 ), D(v 1,v 2 ), and δ(v 1,v 2 ) Theorem 1 (Yang et al. 2005): If δ(t 1,T 2 ) = k, then H ( v, v ) (4q 3) k Theorem 2: for integer vectors, Hv (, v) Dv (, v) Hv (, v) Theorem 3: for tree characteristic vectors, 1 2 Hv ( 1, v2) Dv ( 1, v2) Hv ( 1, v2) k 4q 3 4q 3 4q 3 14

15 Map Trees to Numerical Vectors (1) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Subtree vectors: 15

16 Map Trees to Numerical Vectors (1) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Subtree vectors: 16

17 Map Trees to Numerical Vectors (1) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Subtree vectors: 17

18 Map Trees to Numerical Vectors (1) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Subtree vectors: 18

19 Map Trees to Numerical Vectors (1) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Subtree vectors: 19

20 Map Trees to Numerical Vectors (1) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Subtree vectors: 20

21 Map Trees to Numerical Vectors (1) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Subtree vectors: 21

22 Map Trees to Numerical Vectors (1) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Subtree vectors: 22

23 Map Trees to Numerical Vectors (2) Forest vectors For sequences of program elements Why needed? S1; S2; S3; S4; S0; S5; S2; S3; S6 list list list S6 list S4 list S3 list S3 list S2 S1 S2 S0 S5 23

24 Map Trees to Numerical Vectors (2) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Forest vectors: 24

25 Map Trees to Numerical Vectors (2) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Forest vectors: 25

26 Map Trees to Numerical Vectors (2) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Forest vectors: 26

27 Map Trees to Numerical Vectors (2) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> Forest vectors: 27

28 Map Trees to Numerical Vectors (2) <id, lit, assign_e, incr_e, array_e, cond_e, expr_s, decl, for_s> All vectors: 28

29 Vector Clustering Have millions of generated vectors Need to efficiently cluster close vectors together Hashing Locality-sensitive hashing (LSH) A family of hash functions, s.t., with high probability, Similar vectors hashed to a same hash value Distant vectors hashed to different hash values 29

30 LSH-based Clone Detection (1) Rely on LSH to construct hash tables for vectors h(v i ) v 1 v i v n L S H h(v 1 ) h(v n ) h( ) 30

31 LSH-based Clone Detection (2) Query LSH for vector clusters v 1, v k, v j, v i v n, v m,, v y v 1 Close neighbors of v 1 L S H h(v 1 ) h(v i ) h(v n ) h( ) 31

32 LSH-based Clone Detection (2) Query LSH for vector clusters v 1, v k, v j, v v n, v m, i, v y v i v i Close neighbors of v i L S H h(v 1 ) h(v i ) h(v n ) h( ) 32

33 LSH-based Clone Detection (2) Query LSH for vector clusters v 1, v k, v j, v v n, v m, i, v y v n Close neighbors of v n L S H h(v 1 ) h(v i ) h(v n ) h( ) Post-process to reduce false clones Small clones Overlapping/duplicate clones 33

34 Implementation Modified bison for constructing parse trees 1-level characteristic vectors for trees Group vectors according to their sizes Able to handle size-sensitive clones Less stress on LSH Easy to parallelize 34

35 Evaluation Results Comparison with two state-of-the-art tools Tree-based: CloneDR (for Java) An evaluation version of a commercial tool Token-based: CP-Miner (for C) Evaluation metrics Clone quantity (% of cloned code) Clone quality (% of false clones) Scalability 35

36 Test Programs Application Language # files # lines Linux kernel C 7,988 5,287,090 JDK Java 8,534 2,418,767 GCC C 7,828 1,246,322 PostgreSQL C ,015 Derby Java 1, ,939 Apache C ,287 36

37 Test Programs Application Language # files # lines Linux kernel C 7,988 5,287,090 JDK Java 8,534 2,418,767 GCC C 7,828 1,246,322 PostgreSQL C ,015 Derby Java 1, ,939 Apache C ,287 37

38 Detected More Clones (1) JDK (Deckard vs. CloneDR) Cloned LoC (# / 10^5) Similarity minimum clone size: 50 tokens, stride: 4 CloneDR w/ optimal parameters 38

39 Detected More Clones (2) Linux kernel (Deckard vs. CP-Miner) Gap: # of statement insertions, deletions, and modifications LoC in Clones (# / 10^5) Similarity minimun clone size: 50 tokens; stride: 4 CP-Miner min size: 50 tokens; Gap 0 CP-Miner min size: 50 tokens; Gap 1 39

40 Clone Quality (1) Examined 100 clone clusters Randomly selected from JDK Used parameters Similarity: 1.0 Minimum clone size in tokens: 50 Stride: 4 Among the 100 clusters 93 are clearly real clones The remaining 7 are difficult to assess But they are structurally similar 40

41 Clone Quality (2) if else if (option.equalsignorecase(``basic'')) { bbasictraceon = true; } else if (option.equalsignorecase(``net'')) { bnettraceon = true; } else if (option.equalsignorecase(``security'')) { bsecuritytraceon = true; } else if else if (opt.equals(``-nohelp'')) { nohelp = true; } else if (opt.equals(``-splitindex'')) { splitindex = true; } else if (opt.equals(``-noindex'')) { createindex = false; } else

42 Scalability (1) JDK (Deckard vs. CloneDR) Running Time (min) Similarity minimum clone size: 50 tokens; stride 4 CloneDR w/ optimal parameters 42

43 Scalability (2) Linux kernel (Deckard vs. CP-Miner) Gap: # of statement insertions, deletions, and modifications Running Time (min) minimum clone size: 50 tokens; stride 4 CP-Miner min size: 50 tokens; Gap 0 CP-Miner min size: 50 tokens; Gap Similarity 43

44 Conclusion New algorithm for tree similarity checking Map trees to numerical vectors Cluster similar vectors efficiently Application of the algorithm on clone detection Similarity checking on parse trees Promising results: scalable and accurate 44

45 Thank you! Questions? 45