Regression Test Selection for AspectJ Software

1
Regression Test Selection for AspectJ Software

• Guoqing Xu and Atanas Rountev
• Ohio State University
• ICSE07

2
Outline

• Background and motivation
• Regression test selection
• AspectJ semantics and challenges
• Contributions
• A control-flow representation for AspectJ
  software
• A graph traversal algorithm for test selection
• Experimental evaluation
• Conclusions

3
Regression Test Selection

• Select a safe subset of regression tests
• Harrold et al., OOPSLA 01
• Java interclass graph (JIG) intra- and
  inter-procedural flow of control
• Simultaneous JIG traversal for P and P'

Program P
Execute P and record coverage
JIG edge coverage matrix
Select tests
Dangerous edges in P
Program P
Identify dangerous entities
Program P'
4
Motivation and Challenges

• Aspects can change dramatically the behavior of
  the original code
• Why not select tests based on the woven bytecode?
• The discrepancy between the source code and the
  bytecode can be significant
• Compiler-specific artificial code
• Shown in our experimental study
• A more general question
• What is an appropriate static representation for
  AspectJ software for regression test selection
  and other analyses?

5
AspectJ Semantics

• Join points, pointcuts, and advices
• before, around (with proceed), after
• Shadow
• Textual part of the program executed during the
  time span of a join point
• Dynamic pointcut/advice triggered only when
  certain run-time conditions hold
• if
• Advice precedence rules

6
Running Example
/ before1 / before(Point p, int x)
setterX(p) args(x) throws if () throw
ex / around1 / void around(Point p, int
x) setterX(p) args(x) if(x0)
proceed(p,x) / before2 / before(Point p)
setterX(p) / after1 / after(Point p)
setterX(p)
class Point int x void
setX(int x) this.x x static void
main(String a) Point p new
Point() p.setX(10)
aspect BoundPoint pointcut setterX(Point p)
call(void Point.setX())
target(p) // advices
7
AspectJ Semantic

• try
• before1()
• if (x0) around1()
• else
• before2()
• p.setX()
• catch(Throwable e)
• after1()
• throw e
• after1()

around1() before2() p.setX()
main() p.setX(10)
8
Control-Flow Representation

• AspectJ Inter-module Graph (AJIG)
• JIG-like representation for normal calls
• No join points
• New representation for interactions at join
  points
• Interaction graph
• Multiple advices applicable at a join point
• Model complex interactions among multiple advices
  
• Dynamic advices
• E.g. model the invocation of around1 in the
  example
• void around(Point p, int x) if(x0)

9
Multiple Advice Invocation

• Input a list of advices that statically match a
  shadow
• Sort the list using precedence rules, taking into
  account the shadow
• before1, around1, before2, p.setX, after1
• Build advice nesting tree
• Create a root node, and put every advice under
  root
• Scan the list, build around subtrees advices
  that are invoked within an around advice A are
  the children of A
• Parent-child relationships represent nesting of
  advice time span

10
Advice Nesting Tree

• before1

p.setX
after1
around1
before2
root
before1
around1
after1
p.setX
before2
11
Modeling of proceed

• Important observation
• Advices at one level are invoked by the call to
  proceed in their parent advice
• Introduction of placeholder methods ph_
• A call to a ph_proceed represents a call to
  proceed
• ph_root is called to replace the shadow
• The CFG for a ph_proceed method contains a
  sequence of call/return nodes pairs for all
  advices that are invoked by proceed

12
root
before1
after1
around1
p.setX
before2
13
Handling of After Advices

• Normal subgraph
• after1
• Exceptional subgraph
• after1

Normal subgraph
Exceptional subgraph
14
Handling of Dynamic Advices

• Challenges
• No way to know whether a dynamic advice is
  invoked at run time
• Advices that are nested within a dynamic around
  advice A are still invoked even if A is not
  invoked
• Solutions
• Introduce placeholder decision making nodes
  ph_decision
• Create a ph_decision node before the call node of
  each dynamic advice

15
Example

• Create a ph_decison node
• Create a T edge going from ph_decison to the
  guarded advice
• Create an F edge
• For a non-around advice, link the edge to the
  next call node
• For an around advice, link the edge to the call
  node of its ph_proceed
• Redirect edges

ph_decision
T
F
F
before1
return
ph_decision
T
F
around1
ph_proceed1
return
return
exit
16
Graph Traversal Algorithm

• Edge-level comparison for edges outside the
  interaction graph
• Interaction graph comparison
• Interprocedural traversal
• Compare the calling structure
• Schedule advice bodies for further processing
• Intraprocedural comparison
• Edge-level comparison for advice bodies

17
Inter-procedural traversal
Case 1 if ad1 ! ad2 e is dangerous
18
Inter-procedural traversal
Case 2 e is dangerous
19
Inter-procedural traversal
Case 3 if ad1 ad2 e is dangerous
else T is dangerous 3a if ad1 is
non-around compare(e, e) 3b else
compare(ph, e)
20
Empirical Evaluation
Subject Shadows versions Tests
bean 11 7 42
tracing 32 5 45
Telecom 19 6 41
quicksort 15 3 24
nullcheck 146 4 63
dcm 1103 3 63
lod 359 3 63
21
23.8 more
35 more
22
Study 1 (Cond.)

• Compiler changes names of existing advices when
  adding a new advice
• Compiler inlines an advice when removing some
  control flow paths from it
• Compiler generates try-catch block, when adding
  an afterThrowing advice
• Compiler inserts dynamic residue that performs
  run-time check before dynamic advice
• Conclusion
• These are not program changes
• Such changes prevent JIG-based approach from
  selecting the changes only made in the source.

23
Study 2-Test Suite Reduction
bean
tracing
quicksort
telecom
24
Study 2-Test Suite Reduction
nullcheck
dcm
JIG- 98.8 AJIG- 69.0
loc
25
Conclusions

• Bytecode level analysis does not work for AspectJ
  program
• A source code level control-flow representation
  AJIG
• Multiple advice invocation
• Dynamic advice
• A graph traversal algorithm
• Evaluation
• Our approach outperforms the JIG-based approach

26

• Questions?