Dataflow Testing
Dataflow Testing
PATH SELECTION:
o Select a set of covering paths (c1+c2) using the analogous criteria you used for
structural path testing.
o Select a covering set of paths based on functionally sensible transactions as you
would for control flow graphs.
o Try to find the most tortuous, longest, strangest path from the entry to the exit of
the transaction flow.
PATH SENSITIZATION:
o Most of the normal paths are very easy to sensitize-80% - 95% transaction flow
coverage (c1+c2) is usually easy to achieve.
o The remaining small percentage is often very difficult.
o Sensitization is the act of defining the transaction. If there are sensitization
problems on the easy paths, then bet on either a bug in transaction flows or a
design bug.
PATH INSTRUMENTATION:
o Instrumentation plays a bigger role in transaction flow testing than in unit path
testing.
o The information of the path taken for a given transaction must be kept with that
transaction and can be recorded by a central transaction dispatcher or by the
individual processing modules.
o In some systems, such traces are provided by the operating systems or a running
log.
BUG ASSUMPTION:
The bug assumption for data-flow testing strategies is that control flow is generally
correct and that something has gone wrong with the software so that data objects are not
available when they should be, or silly things are being done to data objects.
o Also, if there is a control-flow problem, we expect it to have symptoms that can
be detected by data-flow analysis.
o Although we'll be doing data-flow testing, we won't be using data flow graphs as
such. Rather, we'll use an ordinary control flow graph annotated to show what
happens to the data objects of interest at the moment.
1 dd :- probably harmless but suspicious. Why define the object twice without an intervening
usage?
2 dk :- probably a bug. Why define the object without using it?
3 du :- the normal case. The object is defined and then used.
4 kd :- normal situation. An object is killed and then redefined.
5 kk :- harmless but probably buggy. Did you want to be sure it was really killed?
6 ku :- a bug. the object doesnot exist.
7 ud :- usually not a bug because the language permits reassignment at almost any time.
8 uk :- normal situation.
9 uu :- normal situation.
In addition to the two letter situations, there are six single letter situations.We will use a leading
dash to mean that nothing of interest (d,k,u) occurs prior to the action noted along the entry-exit
path of interest.
A trailing dash to mean that nothing happens after the point of interest to the exit.
They possible anomalies are:
1 -k :- possibly anomalous because from the entrance to this point on the path, the
variable had not been defined. We are killing a variable that does not exist.
2 -d :- okay. This is just the first definition along this path.
3 -u :- possibly anomalous. Not anomalous if the variable is global and has been
previously defined.
4 k- :- not anomalous. The last thing done on this path was to kill the variable.
5 d- :- possibly anomalous. The variable was defined and not used on this path. But
this could be a global definition.
6 u- :- not anomalous. The variable was used but not killed on this path. Although this
sequence is not anomalous, it signals a frequent kind of bug. If d and k mean
dynamic storage allocation and return respectively, this could be an instance in
which a dynamically allocated object was not returned to the pool after use.
Data flow anomaly model prescribes that an object can be in one of four distinct states:
0. K :- undefined, previously killed, doesnot exist
1. D :- defined but not yet used for anything
2. U :- has been used for computation or in predicate
3. A :- anomalous
These capital letters (K, D, U, A) denote the state of the variable and should not be confused
with the program action, denoted by lower case letters.
Unforgiving Data - Flow Anomaly Flow Graph: Unforgiving model, in which once a variable
becomes anomalous it can never return to a state of grace.
Assume that the variable starts in the K state - that is, it has not been defined or does not exist. If
an attempt is made to use it or to kill it (e.g., say that we're talking about opening, closing, and
using files and that 'killing' means closing), the object's state becomes anomalous (state A) and,
once it is anomalous, no action can return the variable to a working state.
If it is defined (d), it goes into the D, or defined but not yet used, state. If it has been defined (D)
and redefined (d) or killed without use (k), it becomes anomalous, while usage (u) brings it to the
U state. If in U, redefinition (d) brings it to D, u keeps it in U, and k kills it.
Forgiving Data - Flow Anomaly Flow Graph: Forgiving model is an alternate model where
redemption (recover) from the anomalous state is possible
Figure 3.6: Forgiving Data Flow Anomaly State Graph
This graph has three normal and three anomalous states and he considers the kk sequence not to
be anomalous. The difference between this state graph and Figure 3.5 is that redemption is
possible. A proper action from any of the three anomalous states returns the variable to a useful
working state.
The point of showing you this alternative anomaly state graph is to demonstrate that the specifics
of an anomaly depends on such things as language, application, context, or even your frame of
mind. In principle, you must create a new definition of data flow anomaly (e.g., a new state
graph) in each situation. You must at least verify that the anomaly definition behind the theory or
imbedded in a data flow anomaly test tool is appropriate to your situation.
Static analysis is analysis done on source code without actually executing it. For example: source
code syntax error detection is the static analysis result.
Dynamic analysis is done on the fly as the program is being executed and is based on
intermediate values that result from the program's execution. For example: a division by zero
warning is the dynamic result.
If a problem, such as a data flow anomaly, can be detected by static analysis methods, then it
doesn’t belongs in testing - it belongs in the language processor.
There is actually a lot more static analysis for data flow analysis for data flow anomalies going
on in current language processors.
For example, language processors which force variable declarations can detect (-u) and (ku)
anomalies.But still there are many things for which current notions of static analysis are
INADEQUATE.
Why Static Analysis isn't enough? There are many things for which current notions of static
analysis are inadequate. They are:
Dead Variables: Although it is often possible to prove that a variable is dead or alive at a
given point in the program, the general problem is unsolvable.
Arrays: Arrays are problematic in that the array is defined or killed as a single object, but
reference is to specific locations within the array. Array pointers are usually dynamically
calculated, so there's no way to do a static analysis to validate the pointer value. In many
languages, dynamically allocated arrays contain garbage unless explicitly initialized and
therefore, -u anomalies are possible.
Records and Pointers: The array problem and the difficulty with pointers is a special case
of multipart data structures. We have the same problem with records and the pointers to
them. Also, in many applications we create files and their names dynamically and there's no
way to determine, without execution, whether such objects are in the proper state on a given
path or, for that matter, whether they exist at all.
Dynamic Subroutine and Function Names in a Call: subroutine or function name is a
dynamic variable in a call. What is passed, or a combination of subroutine names and data
objects, is constructed on a specific path. There's no way, without executing the path, to
determine whether the call is correct or not.
False Anomalies: Anomalies are specific to paths. Even a "clear bug" such as ku may not be
a bug if the path along which the anomaly exist is unachievable. Such "anomalies" are false
anomalies. Unfortunately, the problem of determining whether a path is or is not achievable
is unsolvable.
Recoverable Anomalies and Alternate State Graphs: What constitutes an anomaly
depends on context, application, and semantics. How does the compiler know which model I
have in mind? It can't because the definition of "anomaly" is not fundamental. The language
processor must have a built-in anomaly definition with which you may or may not (with
good reason) agree.
Concurrency, Interrupts, System Issues: As soon as we get away from the simple single-
task uniprocessor environment and start thinking in terms of systems, most anomaly issues
become vastly more complicated.
How often do we define or create data objects at an interrupt level so that they can be
processed by a lower-priority routine? Interrupts can make the "correct" anomalous and the
"anomalous" correct. True concurrency (as in an MIMD machine) and pseudo concurrency
(as in multiprocessing) systems can do the same to us. Much of integration and system
testing is aimed at detecting data-flow anomalies that cannot be detected in the context of a
single routine.
Although static analysis methods have limits, they are worth using and a continuing trend in
language processor design has been better static analysis methods, especially for data flow
anomaly detection. That's good because it means there's less for us to do as testers and we
have far too much to do as it is.
DATA FLOW MODEL:
The data flow model is based on the program's control flow graph - Don't confuse that with the
program's data flow graph.
Here we annotate each link with symbols (for example, d, k, u, c, and p) or sequences of symbols
(for example, dd, du, ddd) that denote the sequence of data operations on that link with respect to
the variable of interest. Such annotations are called link weights.
The control flow graph structure is same for every variable: it is the weights that change.
Figure 3.9: Control flow graph annotated for X and Y data flows.
INTRODUCTION:
TERMINOLOGY:
1. Definition-Clear Path Segment, with respect to variable X, is a connected
sequence of links such that X is (possibly) defined on the first link and not
redefined or killed on any subsequent link of that path segment. ll paths in Figure
3.9 are definition clear because variables X and Y are defined only on the first
link (1,3) and not thereafter. In Figure 3.10, we have a more complicated
situation. The following path segments are definition-clear: (1,3,4), (1,3,5),
(5,6,7,4), (7,8,9,6,7), (7,8,9,10), (7,8,10), (7,8,10,11). Subpath (1,3,4,5) is not
definition-clear because the variable is defined on (1,3) and again on (4,5). For
practice, try finding all the definition-clear subpaths for this routine (i.e., for all
variables).
2. Loop-Free Path Segment is a path segment for which every node in it is visited
atmost once. For Example, path (4,5,6,7,8,10) in Figure 3.10 is loop free, but path
(10,11,4,5,6,7,8,10,11,12) is not because nodes 10 and 11 are each visited twice.
3. Simple path segment is a path segment in which at most one node is visited
twice. For example, in Figure 3.10, (7,4,5,6,7) is a simple path segment. A simple
path segment is either loop-free or if there is a loop, only one node is involved.
4. A du path from node i to k is a path segment such that if the last link has a
computational use of X, then the path is simple and definition-clear; if the
penultimate (last but one) node is j - that is, the path is (i,p,q,...,r,s,t,j,k) and link
(j,k) has a predicate use - then the path from i to j is both loop-free and definition-
clear.
STRATEGIES: The structural test strategies discussed below are based on the program's
control flow graph. They differ in the extent to which predicate uses and/or computational uses
of variables are included in the test set. Various types of data flow testing strategies in decreasing
order of their effectiveness are:
All - du Paths (ADUP): The all-du-paths (ADUP) strategy is the strongest data-flow testing
strategy discussed here. It requires that every du path from every definition of every variable to
every some test.
For variable X and Y:In Figure 3.9, because variables X and Y are used only on link (1,3), any test
that starts at the entry satisfies this criterion (for variables X and Y, but not for all variables as
required by the strategy).
For variable Z: The situation for variable Z (Figure 3.10) is more complicated because the
variable is redefined in many places. For the definition on link (1,3) we must exercise paths that
include subpaths (1,3,4) and (1,3,5). The definition on link (4,5) is covered by any path that
includes (5,6), such as subpath (1,3,4,5,6, ...). The (5,6) definition requires paths that include
subpaths (5,6,7,4) and (5,6,7,8).
For variable V: Variable V (Figure 3.11) is defined only once on link (1,3). Because V has a
predicate use at node 12 and the subsequent path to the end must be forced for both directions at
node 12, the all-du-paths strategy for this variable requires that we exercise all loop-free
entry/exit paths and at least one path that includes the loop caused by (11,4).
Note that we must test paths that include both subpaths (3,4,5) and (3,5) even though neither of
these has V definitions. They must be included because they provide alternate du paths to the V
use on link (5,6). Although (7,4) is not used in the test set for variable V, it will be included in
the test set that covers the predicate uses of array variable V() and U.
The all-du-paths strategy is a strong criterion, but it does not take as many tests as it might seem
at first because any one test simultaneously satisfies the criterion for several definitions and uses
of several different variables.
All Uses Startegy (AU):The all uses strategy is that at least one definition clear path from every
definition of every variable to every use of that definition be exercised under some test.
Just as we reduced our ambitions by stepping down from all paths (P) to branch coverage (C2),
say, we can reduce the number of test cases by asking that the test set should include at least one
path segment from every definition to every use that can be reached by that definition.
For variable V: In Figure 3.11, ADUP requires that we include subpaths (3,4,5) and (3,5) in
some test because subsequent uses of V, such as on link (5,6), can be reached by either
alternative. In AU either (3,4,5) or (3,5) can be used to start paths, but we don't have to use both.
Similarly, we can skip the (8,10) link if we've included the (8,9,10) subpath.
Note the hole. We must include (8,9,10) in some test cases because that's the only way to reach
the c use at link (9,10) - but suppose our bug for variable V is on link (8,10) after all? Find a
covering set of paths under AU for Figure 3.11.
All p-uses/some c-uses strategy (APU+C) : For every variable and every definition of that
variable, include at least one definition free path from the definition to every predicate use; if
there are definitions of the variables that are not covered by the above prescription, then add
computational use test cases as required to cover every definition.
For variable Z:In Figure 3.10, for APU+C we can select paths that all take the upper link (12,13)
and therefore we do not cover the c-use of Z: but that's okay according to the strategy's definition
because every definition is covered.
Links (1,3), (4,5), (5,6), and (7,8) must be included because they contain definitions for variable
Z. Links (3,4), (3,5), (8,9), (8,10), (9,6), and (9,10) must be included because they contain
predicate uses of Z. Find a covering set of test cases under APU+C for all variables in this
example - it only takes two tests.
All c-uses/some p-uses strategy (ACU+P) : The all c-uses/some p-uses strategy (ACU+P) is to
first ensure coverage by computational use cases and if any definition is not covered by the
previously selected paths, add such predicate use cases as are needed to assure that every
definition is included in some test.
For variable Z: In Figure 3.10, ACU+P coverage is achieved for Z by path (1,3,4,5,6,7,8,10,
11,12,13[lower], 2), but the predicate uses of several definitions are not covered. Specifically,
the (1,3) definition is not covered for the (3,5) p-use, the (7,8) definition is not covered for the
(8,9), (9,6) and (9, 10) p-uses.
The above examples imply that APU+C is stronger than branch coverage but ACU+P may be
weaker than, or incomparable to, branch coverage.
All Definitions Strategy (AD) : The all definitions strategy asks only every definition of every
variable be covered by atleast one use of that variable, be that use a computational use or a
predicate use.
For variable Z: Path (1,3,4,5,6,7,8, . . .) satisfies this criterion for variable Z, whereas any entry/exit path
satisfies it for variable V.
From the definition of this strategy we would expect it to be weaker than both ACU+P and
APU+C.
1. All Predicate Uses (APU), All Computational Uses (ACU) Strategies : The all predicate
uses strategy is derived from APU+C strategy by dropping the requirement that we include a c-
use for the variable if there are no p-uses for the variable. The all computational uses strategy is
derived from ACU+P strategy by dropping the requirement that we include a p-use for the
variable if there are no c-uses for the variable.
It is intuitively obvious that ACU should be weaker than ACU+P and that APU should be
weaker than APU+C.
ORDERING THE STRATEGIES:
Figure 3.12compares path-flow and data-flow testing strategies. The arrows denote that the
strategy at the arrow's tail is stronger than the strategy at the arrow's head
o The right-hand side of this graph, along the path from "all paths" to "all
statements" is the more interesting hierarchy for practical applications.
o Note that although ACU+P is stronger than ACU, both are incomparable to the
predicate-biased strategies. Note also that "all definitions" is not comparable to
ACU or APU.