 DATA FLOW TESTING:

o Data flow testing is the name given to a family of test strategies based on
selecting paths through the program's control flow in order to explore
sequences of events related to the status of data objects.
o For example, pick enough paths to assure that every data object has been
initialized prior to use or that all defined objects have been used for
something.
o Motivation:

it is our belief that, just as one would not feel confident about a
program without executing every statement in it as part of some
test, one should not feel confident about a program without having
seen the effect of using the value produced by each and every
computation.

 DATA FLOW MACHINES:

o There are two types of data flow machines with different architectures. (1)
Von Neumann machines (2) Multi-instruction, multi-data machines
(MIMD).
o Von Neumann Machine Architecture:
 Most computers today are von-neumann machines.
 This architecture features interchangeable storage of instructions
and data in the same memory units.
 The Von Neumann machine Architecture executes one instruction
at a time in the following, micro instruction sequence:
1. Fetch instruction from memory
2. Interpret instruction
3. Fetch operands
4. Process or Execute
5. Store result
6. Increment program counter
7. GOTO 1
o Multi-instruction, Multi-data machines (MIMD) Architecture:
 These machines can fetch several instructions and objects in
parallel.
 They can also do arithmetic and logical operations simultaneously
on different data objects.
 The decision of how to sequence them depends on the compiler.
 BUG ASSUMPTION:
o 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
flowgraphs as such. Rather, we'll use an ordinary control flowgraph
annotated to show what happens to the data objects of interest at the moment.
 DATA FLOW GRAPHS:
o The data flow graph is a graph consisting of nodes and directed links.

Figure 3.4: Example of a data flow graph

o We will use an control graph to show what happens to data objects of interest
at that moment.
o Our objective is to expose deviations between the data flows we have and the
data flows we want.
o Data Object State and Usage:
 Data Objects can be created, killed and used.
 They can be used in two distinct ways: (1) In a Calculation (2) As
a part of a Control Flow Predicate.
 The following symbols denote these possibilities:
1. Defined: d - defined, created, initialized etc
2. Killed or undefined: k - killed, undefined, released etc
 3. Usage: u - used for something (c - used in Calculations,
p - used in a predicate)
 1. Defined (d):
 An object is defined explicitly when it appears in a data
declaration.
 Or implicitly when it appears on the left hand side of the
assignment.
 It is also to be used to mean that a file has been opened.
 A dynamically allocated object has been allocated.
 Something is pushed on to the stack.
 A record written.
 2. Killed or Undefined (k):
 An object is killed on undefined when it is released or
otherwise made unavailable.
 When its contents are no longer known with certitude
(with aboslute certainity / perfectness).
 Release of dynamically allocated objects back to the
availability pool.
 Return of records.
 The old top of the stack after it is popped.
 An assignment statement can kill and redefine
immediately. For example, if A had been previously
defined and we do a new assignment such as A : = 17,
we have killed A's previous value and redefined A
 3. Usage (u):
 A variable is used for computation (c) when it appears on
the right hand side of an assignment statement.
 A file record is read or written.
 It is used in a Predicate (p) when it appears directly in a
predicate.
 DATA FLOW ANOMALIES:
o An anomaly is denoted by a two-character sequence of actions.
o For example, ku means that the object is kille
o d and then used, where as dd means that the object is defined twice without
an intervening usage.
o What is an anomaly is depend on the application.
o There are nine possible two-letter combinations for d, k and u. some are
bugs, some are suspicious, and some are okay.
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.
o In addition to the two letter situations, there are six single letter situations.
o 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.
o A trailing dash to mean that nothing happens after the point of interest to the
exit.
o They possible anomalies are:
0. -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.
1. -d :- okay. This is just the first definition along this path.
2. -u :- possibly anomalous. Not anomalous if the variable is global
and has been previously defined.
3. k- :- not anomalous. The last thing done on this path was to kill the
variable.
4. d- :- possibly anomalous. The variable was defined and not used
on this path. But this could be a global definition.
5. 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 STATE GRAPH:
o 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
o 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.
o Unforgiving Data - Flow Anomaly Flow Graph:Unforgiving model, in
which once a variable becomes anomalous it can never return to a state of
grace.
 Figure 3.5: Unforgiving Data Flow Anomaly State Graph

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.

o 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 Vs DYNAMIC ANOMALY DETECTION:

o Static analysis is analysis done on source code without actually executing it.
For example: source code syntax error detection is the static analysis result.
o 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.
o If a problem, such as a data flow anomaly, can be detected by static analysis
methods, then it doesnot belongs in testing - it belongs in the language
processor.
o There is actually a lot more static analysis for data flow analysis for data
flow anomalies going on in current language processors.
o For example, language processors which force variable declarations can
detect (-u) and (ku) anomalies.
o But still there are many things for which current notions of static analysis are
INADEQUATE.
o 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 pseudoconcurrency (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.
o 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:
o The data flow model is based on the program's control flow graph - Don't
confuse that with the program's data flowgraph..
o Here we annotate each link with symbols (for example, d, k, u, c, 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.
o The control flow graph structure is same for every variable: it is the weights
that change.
o Components of the model:
0. To every statement there is a node, whose name is unique. Every
node has at least one outlink and at least one inlink except for exit
nodes and entry nodes.
1. Exit nodes are dummy nodes placed at the outgoing arrowheads of
exit statements (e.g., END, RETURN), to complete the graph.
Similarly, entry nodes are dummy nodes placed at entry statements
(e.g., BEGIN) for the same reason.
2. The outlink of simple statements (statements with only one
outlink) are weighted by the proper sequence of data-flow actions
for that statement. Note that the sequence can consist of more than
one letter. For example, the assignment statement A:= A + B in
most languages is weighted by cd or possibly ckd for variable A.
Languages that permit multiple simultaneous assignments and/or
compound statements can have anomalies within the statement.
The sequence must correspond to the order in which the object
code will be executed for that variable.
3. Predicate nodes (e.g., IF-THEN-ELSE, DO WHILE, CASE) are
weighted with the p - use(s) on every outlink, appropriate to that
outlink.
4. Every sequence of simple statements (e.g., a sequence of nodes
with one inlink and one outlink) can be replaced by a pair of nodes
that has, as weights on the link between them, the concatenation of
link weights.
5. If there are several data-flow actions on a given link for a given
variable, then the weight of the link is denoted by the sequence of
actions on that link for that variable.
6. Conversely, a link with several data-flow actions on it can be
replaced by a succession of equivalent links, each of which has at
most one data-flow action for any variable.
o Let us consider the example:

Figure 3.7: Program Example (PDL)

Figure 3.8: Unannotated flowgraph for example program in

Figure 3.7
Figure 3.9: Control flowgraph annotated for X and Y data flows.

Figure 3.10: Control flowgraph annotated for Z data flow.

 Figure 3.11: Control flowgraph annotated for V data flow.

STRATEGIES OF DATA FLOW TESTING:

o Data Flow Testing Strategies are structural strategies.

o In contrast to the path-testing strategies, data-flow strategies take into
account what happens to data objects on the links in addition to the raw
connectivity of the graph.
o In other words, data flow strategies require data-flow link weights
(d,k,u,c,p).
o Data Flow Testing Strategies are based on selecting test path segments (also
called sub paths) that satisfy some characteristic of data flows for all data
objects.
o For example, all subpaths that contain a d (or u, k, du, dk).
o A strategy X is stronger than another strategy Y if all test cases produced
under Y are included in those produced under X - conversely for weaker.
 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.
5. 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 flowgraph. 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:

0. 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 use of that definition be exercised
under 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.

1. 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.

2. 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.

For variable V:In Figure 3.11, APU+C is achieved for V by

(1,3,5,6,7,8,10,11,4,5,6,7,8,10,11,12[upper], 13,2) and
(1,3,5,6,7,8,10,11,12[lower], 13,2). Note that the c-use at (9,10) need not be
included under the APU+C criterion.

3. 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.

4. 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.

5. 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:

o Figure 3.12 compares 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.
 Figure 3.12: Relative Strength of Structural Test Strategies.

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.
 SLICING AND DICING:
o A (static) program slice is a part of a program (e.g., a selected set of
statements) defined with respect to a given variable X (where X is a simple
variable or a data vector) and a statement i: it is the set of all statements that
could (potentially, under static analysis) affect the value of X at statement i -
where the influence of a faulty statement could result from an improper
computational use or predicate use of some other variables at prior
statements.
o If X is incorrect at statement i, it follows that the bug must be in the program
slice for X with respect to i
o A program dice is a part of a slice in which all statements which are known
to be correct have been removed.
 o In other words, a dice is obtained from a slice by incorporating information
obtained through testing or experiment (e.g., debugging).
o The debugger first limits her scope to those prior statements that could have
caused the faulty value at statement i (the slice) and then eliminates from
further consideration those statements that testing has shown to be correct.
o Debugging can be modeled as an iterative procedure in which slices are
further refined by dicing, where the dicing information is obtained from ad
hoc tests aimed primarily at eliminating possibilities. Debugging ends when
the dice has been reduced to the one faulty statement.
o Dynamic slicing is a refinement of static slicing in which only statements on
achievable paths to the statement in question are included.

SUMMARY:

 The methods discussed for path testing of units and programs can be applied with
suitable interpretation to functional testing based on transaction flows.
 The biggest problem and the biggest payoff may be getting the transaction flows in the
first place.
 Full coverage (C1 + C2) is required for all flows, but most bugs will be found on the
strange, meaningless, weird paths.
 Transaction-flow control may be implemented by means of an undeclared and
unrecognized internal language.
 The practice of attempting to design tests based on transaction-flow representation of
requirements and discussing those attempts with the designer can unearth more bugs
than any tests you run.
 Data are as important as code and will become more important.
 Data integrity is as important as code integrity. Just as common sense dictates that all
statements and branches be exercised on under test, all data definitions and subsequent
uses must similarly be tested.
 What constitutes a data flow anomaly is peculiar to the application. Be sure to have a
clear concept of data flow anomalies in your situation.
 Use all available tools to detect those anomalies that can be detected statically. Let the
extent and excellence of static data-flow anomaly detection be as important a criterion
in selecting a language processor as produced object code efficiency and compilation
speed. Use the slower compiler that gives you slower object code if it can detect more
anomalies. You can always recompile the unit after it has been debugged.
 The data-flow testing strategies span the gap between all paths and branch testing. Of
the various available strategies, AU probably has the best payoff for the money. It
seems to be no worse than twice the number of test cases required for branch testing,
but the resulting code is much more reliable. AU is not too difficult to do without
supporting tools, but use the tools as they become available.
 Don't restrict your notion of data-flow anomaly to the obvious. The symbols d, k, u,
and the associated anomalies, can be interpreted (with profit) in terms of file opening
and closing, resource management, and other applications.

TOP