Automatic Program Correction

Report
Automatic Program Correction
Anton Akhi
Friday, July 08, 2011
Plan
• Introduction
• Survey of existing tools
2
Why do we need automatic program
correction?
• Even if bug has been found it is still
programmer’s job to think of fix
• Allows to fix bug automatically or provide
high-quality fix suggestions
3
Isn’t it a magic?
• Tools require failing and passing runs
• Oracle to check if test passes
• Program with bug is not far away from the
right one
4
Existing tools
• Genetic programming:
– Automatic Program Repair with Evolutionary Computation
– A Novel Co-evolutionary Approach to Automatic Software
Bug Fixing
• Machine learning:
– BugFix
• Contract usage:
– Automated Debugging using Path-Based Weakest
Preconditions
– AutoFix-E
– AutoFix-E2
5
Automatic Program Repair with
Evolutionary Computation
• W. Weimer, S. Forrest, C. Le Goues, T. Nguyen
• Usage of GP to fix bugs:
– Individuals
– Genetic Operators
– Fitness function
6
Genetic programming: individuals
• Individual is represented as an abstract syntax
tree and weighted path
• Weighted path is a list of statements visited on
negative test case with weight of every
statement:
– 1 if not visited by positive tests
– 0.1 if visited by positive tests
7
Genetic programming: operators
• Mutation:
– statement on a weighted pass is considered for
mutation with probability proportional to its
weight
– deletion, insertion, swap
• Crossover:
– exchange of subtrees chosen at random between
two individuals
8
Genetic programming: fitness function
• Weighted sum of test cases passed
• Weight of a negative test is at least as much as
positive test
9
Minimizing changes
• Trim unnecessary edits
• One-minimal subset of changes is a subset
such that without any of the changes program
will stop passing all the tests
• Use delta debugging to compute one-minimal
subset of changes
10
Automatic Program Repair with
Evolutionary Computation:
Conclusions
• Needs:
– negative and positive test cases
– oracle
– fault localization
• Uses:
– genetic programming
– Delta debugging
• Method has a lot of criticism
11
A Novel Co-evolutionary Approach to
Automatic Software Bug Fixing
•
•
•
•
•
A. Arcuri and X. Yao
Genetic Programming
Distance Functions
Search Based Software Testing
Co-evolution
12
Genetic Programming
• Genetic Program consists of primitives
• Program is represented in a tree form
• Fitness function to be minimized:
N g 
E g 
f g  

  d P t, g t 
N g   1 E g   1 tTi
• N g  is a number of nodes in a program
• E g  is a number of raised exceptions
• d P t, g t  is a special distance function
13
Distance Functions
• Works fine
with numbers
and boolean
expressions
• Predicates
involving 
and  can be
handled only
in cases small
Q
14
Search Based Software Testing
• Find tests that make evolutionary programs
fail
• Fitness function for test case to be maximized:
f t   d P t, g t 
gGi
15
Co-evolution
• Competitive co-evolution
• First generation: copies of buggy program and
unit tests
• Mutations, crossover, replacement by the
original program
• Penalty for short programs
16
A Novel Co-evolutionary Approach to
Automatic Software Bug Fixing:
Conclusions
• Needs:
– starting set of unit tests
– oracle
• Uses:
– genetic programming
– co-evolution
– search based software testing
• Some bugs are too difficult to solve
• Bug that is difficult to be fixed by a human
might be very easy for program
17
BugFix: A Learning-Based Tool to Assist
Developers in Fixing Bugs
•
•
•
•
•
D. Jeffrey, M. Feng, N. Gupta, R. Gupta
Association rule learning
Interesting value mapping pairs (IVMP)
Situation descriptors
Knowledgebase of rules
18
Association rule learning
• Association rule learning is a popular method for
discovering the relationship between variables in
database
• Association rule X  Y where X, Y are sets of
attributes and X Y   means that if the items
in set X are present then it is probable that the
items in set Y are also present
• The confidence of the rule is
conf  X  Y   supp X  Y  supp X 
where supp(X) is the fraction of transactions
containing X
19
Interesting Value Mapping Pairs
• An Interesting Value Mapping Pair (IVMP) is a pair
of value mappings (original, alternate) associated
with a particular statement instance in a failing
run, such that: (1) original is the original value
mapping used by the failing run at that instance;
and (2) alternate is an alternate (different) value
mapping such that if the values in original are
replaced by the values in alternate at that
instance during re-execution of the failing run,
then the incorrect output of the failing run
becomes correct
20
Situation descriptors
• Statement structure situation descriptors
• IVMP pattern situation descriptors
• Value pattern situation descriptors
21
Statement structure situation
descriptors
• Unordered tokens comprising the statement
22
Statement structure situation
descriptors: examples
23
IVMP pattern situation descriptors
• Consider pattern to occur when corresponding
values in the IVMPs compare to each other in
the same way across all IVMPs at a statement
• Compare in terms less than, greater than, or
equal to
• Look at the pairs of values:
– within original sets of values
– within alternate sets of values
– between corresponding values in sets
24
IVMP pattern situation descriptors:
examples
25
Value pattern situation descriptors
• Similar to IVMP patterns
26
Knowledgebase of rules
• Database of bug-fix scenarios
• Initially created through training data of
known debugging situations
27
How it works
• Identify rules to consider
– rules in which debugging situation is subset of the
current debugging situation
• Sort rules by confidence values
• Report prioritized bug-fix descriptions
• Learn from current debugging situation and
the corresponding bug fix
28
BugFix: A Learning-Based Tool to Assist
Developers in Fixing Bugs: Conclusions
• BugFix assists in fixing bugs by producing a list
of bug-fix suggestions
• Tool learns through new situations
• Is not very good with new and logically
difficult bugs
29
Automated Debugging using
Path-Based Weakest Preconditions
•
•
•
•
•
•
•
H. He, N. Gupta
Representation of an error trace
Path-based weakest precondition
Hypothesized program state
Actual program state
Detection of evidence
Location and modification of likely errorneous
statement
30
Representation of an error trace
• i, j is an instance of an executed statement in an
error trace; i is an execution point and j is the line
number of statement
• Branch predicates:
– atomic predicate: (expr relop const)
– compound predicates are in disjunctive normal form:
E  e0    en where ei  g0  g n and gi is an
atomic predicate
• Precondition and postcondition for failing run are
transformed in DNF:
– replace  and  by disjunction and conjunction
31
Path-based weakest precondition
• pwp(T, R) is the set of all states such that an
execution of function F that flows execution
trace T begun in any of them is guaranteed to
terminate is state satisfying R
• pwp( x  a, R)  Rxa where xa means
substituting every occurrence of x in R with a
• pwp( B, R)  R where B is branch predicate
• pwp(C; D, R)  pwp(C, pwp( D, R))
32
Hypothesized program state
• Let Ri  pwpT i,n , R where T i,n is a trace from
point i to the end of trace and R is
postcondition
• Ri defines the set of hypothesized program
states at an execution point i
• Ri  pwpStmti , Ri 1  and Rn  pwpStmtn , R
33
Actual program state
• Represented by predicates in DNF which are
true for given input
F
Q
• Consists of forward program states and i
B
Q
backward program states i
F
B
Q

Q

Q
• i i
i
34
Forward and backward program states
• Forward program states are defined as:
–
–
–
–
Q1F  positive conjunctions in precondition
QiF  QiF1  Killi1   Geni1
F
Qbottom
 QnF  Killn  Genn


Killi and Geni are sets of predicates killed by and
derived from statement Stm ti
• Backward program states are defined as:
– Q  pwpStmt ,Q  if Stm ti is an assignment statement
B
i
i
B
i 1
– QiB  Stmti  QiB1 if Stm ti is a branch predicate
B
 
– Qbottom
35
Detection of evidence
• A is less restrictive than B if A  B is false
• Evidence at point i is situation when Qi is less
restrictive than Ri
• Two types of evidence:
– Explicit
• Type I
• Type II
– Implicit
36
Types of evidence
• Explicit
– Type I: if at point i in Ri appears negative r in form
(0 relop const) then r forms Eexplicitany, r, i  an
explicit evidence of Type I
– Type II: let q : expr q relop const 1 an atomic predicate in Qi
and r : expr r relop const 2 a negative atomic predicate in Ri
If exprq  exprr then q and r form Eexplicitq, r, i  an explicit
evidence of Type II
• Implicit: negative predicate r in R1 that is not
present in an explicit evidence
37
Location and modification of likely
erroneous statement
• Use transitivity and equality to deduce new
predicates. New states are Qi* and Ri*
• Explicit Type I
– If = then match r to 0=0. Consider every
assignment statement between i and bottom of
the trace as a possible candidate for modification
– Let rk from Rk* be a corresponding predicate to r,
Stmtk : lhs  rhs and e  lhs  rhs  0 and c  rk 1
– Goal is to make rk  0  0, so pwplhs  rhs, rk 1   0  0
– If rk 1  lhs  rhs  0 then problem is solved
38
Location and modification of likely
erroneous statement: Explicit Type I
• Consider e and c as a set of strings of characters
• Let de and d c be difference between e and c and
between c and e correspondingly
• If de appears in rhs than replace it with d c
• If de appears in lhs than replace it with d c only if d c
is a single variable
• If none of the above works than select r as e and
try every q from Qi* as c
39
Location and modification of likely
erroneous statement: Explicit Type II
• Consider Eexplicitq, r, i 
• Either q or r could be in error
• Change the form of r to q at i
– Same manner as above
• Change the form of q to r at i
– Change original branch predicate from which q
may be derived or some assignment statement
– Change of an assignment statement does not
change relop in q
40
Location and modification of likely
erroneous statement: Implicit
• If there is a loop in the trace, which contributes
some constraints on R1* , and missed constraints
have similarity with constraints added by the
loop then try to derive the possible missing
iterations in the loop
• Try to match negative r from R1* to some q from Q1*
41
Automated Debugging using
Path-Based Weakest Preconditions:
Conclusions
• Uses contracts
• Can handle only one error
• Cannot handle loops well
42
Automated Fixing of Programs with
Contracts
• Yi Wei, Yu Pei, C.A. Furia, L.S. Silva, S.
Buchholz, B. Meyer, A. Zeller
• Assessing Object State
• Fault Analysis
• Behavioral Models
• Generating Candidate Fixes
• Linearly Constrained Assertions
• Validating and Ranking Fixes
43
Assessing Object State
• Argument-less Boolean Queries
– Absolutely describe state
– Seldom preconditions
– Widely used in Eiffel contracts
• Complex Predicates
– Boolean queries are often correlated
– Implication expresses correlation
– Mine the contracts
– Mutate implications
44
Fault Analysis
• Find state invariants: passing and failing states
– sets of predicates that hold during all
passing and failing runs respectively
• Fault profile contains all predicates that hold
in the passing run but not in the failing run
• Find the strongest predicate that implies the
negation of violated assertion
45
Behavioral models
• Finite state automaton
– States are predicates that hold
– Transitions are routines
• Automaton is built based on test runs
• Determine a sequence of routine calls that
change the object state appropriately
• A snippet for a set of predicates is any
sequence of routines that drive object from a
state where none of predicates hold to one
where all of them hold
46
Generating Candidate Fixes
• Fix Schemas
• snippet is a sequence of routine calls
• old_stmt – some statements in the original
program
• fail:
– not p
– not p1 and not p2 and
– notviolated_ clause
47
Linearly Constrained Assertions
• Determine what is variable and what is
constant
• Assign weights:
– Arguments in precondition receive lower weights
– In assertion weight is inversely proportional to the
number of occurrences
– Identifiers that routine can assign receive less
weight
48
Linearly Constrained Assertions:
Generating fixes
• Select a value for variable that satisfies the
constraint
– Look for extremal values
• Plug the value into a fix schema
– if not constraint then new_stmt else old_stmt end
49
Validating and Ranking Fixes
• Candidate is valid if it passes all the tests
• Two metrics for ranking:
– Dynamic: estimates the difference in runtime
behavior between the fix and the original based
on state distance
– Static: OS  5  SN  2.5  BF
• OS: 0 for schemas (a) and (b) and number of
statements in old_stmt for (c) and (d)
• SN: number of statements in snippet
• BF: number of branches to reach old_stmt from the
point of injection of the instantiated fix schema
50
Automated Fixing of Programs with
Contracts: Conclusions
• Uses contracts, passing and failing tests to
deduce and suggest bug fixes
• Successfully proposed fixes for 16 out of 42
found bugs in EiffelBase library
51
Evidence-Based Automated Program
Fixing
•
•
•
•
•
•
•
Yu Pei, Yi Wei, C.A. Furia, M. Nordio, B. Meyer
Predicates, Expressions, and States
Static Analysis
Dynamic Analysis
Fixing Actions
Fix Candidate Generation
Validation of Candidates
52
Predicates, Expressions
• E r is a set of all non-constant expressions in
routine r
• Pr is a set of boolean predicates:
– Boolean expressions: every b  Er
– Voidness checks: e  Void for every e  Er
– Integer comparisons:e ~ e for every e  Er and
every e  Er \ e 0 and ~ in , , 
– Complements  p for every p  Pr
53
States
• State Components
– l , p, v - a triple where v is a value of predicate p
for some test case t which riches l
– comp(T) denotes all the triples l , p, v defined by
the tests in the set t  T
54
Static Analysis
• sub(e) is the set of all sub-expressions of e
• Expression proximity eproxe1 , e2   sube1   sube2 
• Expression dependence between predicate p  Pr
eprox p, c 


and a contract clause c edep p, c  maxeprox , c |   P 
r
• Control distance cdistl1 ,l2  is the length of the
shortest directed path from l1 to l2 on the
control-flow graph
• Control dependence
cdistl , j 
cdepl , j   1 
maxcdist , j  |   r and  j
55
Dynamic Analysis
•  t  is a score for tests
i


•  t   for i-th failing test and  t    i for i-th
passing test; 0   ,   1
• The evidence provided by each test case:
dyn l, p, v     u | u  Fr j,c   v | v  Pr 
56
Combining Static and Dynamic Analysis
• Combined evidence score:
fixme l , p, v 
3
edep p, c   cdepl , j   dyn l , p, v
1
1
1
57
Fixing Actions
• A component l , p, v with a high evidence
score induces a number of possible actions:
– Derived Expressions
– Expression Modification
– Expression Replacement
58
Fix Candidate Generation
• Candidates are generated in a way similar to
previous method
• Candidate is valid if it passes all the tests
59
Evidence-Based Automated Program
Fixing: Conclusions
• Uses contracts and sets of passing and failing
tests
• Combines static and dynamic approaches
60
Conclusion
• Automated Bug Fixing is real!
• Some bugs are still too difficult to fix or even
localize
• All approaches need some kind of oracle;
sometimes contracts are the oracle
61
References
• W. Weimer, S. Forrest, C. Le Goues, T. Nguyen,
Automatic Program Repair with Evolutionary
Computation
• A. Arcuri and X. Yao, A Novel Co-evolutionary Approach
to Automatic Software Bug Fixing
• D. Jeffrey, M. Feng, N. Gupta, R. Gupta, BugFix: A
Learning-Based Tool to Assist Developers in Fixing Bugs
• H. He, N. Gupta, Automated Debugging using
Path-Based Weakest Preconditions
• Yi Wei, Yu Pei, C.A. Furia, L.S. Silva, S. Buchholz, B.
Meyer, A. Zeller, Automated Fixing of Programs with
Contracts
• Yu Pei, Yi Wei, C.A. Furia, M. Nordio, B. Meyer,
Evidence-Based Automated Program Fixing
62

similar documents