COUNTEREXAMPLE DRIVEN REFINEMENT FOR ABSTRACT INTERPRETATION
BACKGROUND
[0001] Abstract interpretation is a theory for approximating the semantics of systems, such as computer programs and industrial designs, based on monotonic functions over ordered sets, especially lattices. Abstract interpretation can be viewed as a partial execution of a computer program that gains information about its semantics (e.g., control structure and flow of information) without performing all the calculations of the program. Abstract interpretation computes fixpoints for programs that each represent a state of the abstract interpretation of the program at some point or step of interpretation. Abstract interpretation can be used to determine whether a program satisfies a condition (e.g., the value of variable B is never greater than 100). If all fixpoints computed by an abstract interpretation of the program satisfy the condition, then all concrete behaviors of the program satisfy the condition. If a fixpoint, however, does not satisfy the condition, then there are two possibilities. First, the program does not satisfy the condition, and a "true error" in the program has been found. Second, the program satisfies the condition but the abstract interpretation was not precise enough to verify it, and a "false error" in the program has been found. Abstract interpretations represent programs with varying degrees of precision to accommodate programs with infinite domains or to efficiently accommodate very large programs. If the precision is too low, the abstract interpretation will, however, detect many false errors, which reduces its usability.
[0002] Predicate abstraction is a form of abstract interpretation in which concrete states are abstracted by their valuation to a finite set of predicates. Several tools are available that use predicate abstraction to verify finite state interface protocols on programs. To reduce false errors, these tools typically analyze an abstract counterexample to check whether the counterexample is feasible in the concrete program. A counterexample indicates that the condition of the program is not satisfied. If the counterexample is feasible, the toots add more predicates to
improve the precision of the predicate abstraction. This process, referred to as "counterexample driven refinement," continues iteratively until (1) the condition is proved, (2) a true error is found, or (3) resources (e.g., time or memory) are exhausted.
[0003] Abstract interpretations generally operate over mathematical lattices and compute the semantics of programs as fixpoints. Such fixpoint computations may not converge, however, if the lattice has infinite ascending chains. To ensure convergence of fixpoints, abstract interpretations use a "widening" technique to over-approximate the semantics. The widening operator V has the property that for all states x and y the result *Vv is greater than both x and y. Furthermore, widening guarantees convergence of fixpoint computation in that in any infinite increasing sequence x0, x1, x2,K , the sequence y0, y1, y2,K, given by y0 = xo and
yi+1 = yi▼(yiUi+1) is guaranteed to converge. SUMMARY
[0004] A method and system for verifying an abstract interpretation of a system using counterexample driven refinement is provided. The refinement system abstractly represents a target system as a transition system with an initial state and state transitions and a specification that the target system is to satisfy. The refinement system iteratively performs steps of the abstract interpretation starting with the initial state to generate resulting states. The refinement system then applies a widening operator to the resulting states to establish over-approximated states of the step. After the states of a step are established, the refinement system determines whether the states of the step include an error state and thus will not satisfy the specification. If the states include an error state, the refinement system determines whether the error is a false error or a true error. To make this determination, the refinement system identifies which of the previous steps established a state as a result of the over-approximation of the widening operator that led to the error state. If the refinement system does not identify such a step, then the error is a true error. If the refinement system does identify such a step, the refinement system then applies a more precise operator at that step to establish new states for that step that do not include the state that led to the error state. The
refinement system then re-performs the next step using the new states to establish states and continues this process for subsequent steps using the widening operator until the states converge.
BRIEF DESCRIPTION OF THE DRAWINGS
[0005] Figure 1 is a block diagram that illustrates components of the refinement system in one embodiment.
[0006] Figure 2 is a flow diagram that illustrates the processing of the check abstract interpretation component of the refinement system in one embodiment.
[0007] Figure 3 is a flow diagram that illustrates the processing of the identify source component of the refinement system in one embodiment.
[0008] Figure 4 is a flow diagram that illustrates the processing of an abstract refine component of the refinement system in one embodiment.
[0009] Figure 5 is a flow diagram that illustrates the processing of the abstract fixpoint component of the refinement system in one embodiment.
[0010] Figure 6 is a flow diagram that illustrates the processing of a refine component of the refinement system in one embodiment.
DETAILED DESCRIPTION
[0011] A method and system for verifying an abstract interpretation of a system using counterexample driven refinement is provided. In one embodiment, the refinement system abstractly represents a target system, such as a computer program, as a transition system with an initial state and state transitions, and a specification that the target system is to satisfy. The state transitions specify possible transitions from one state to another state. The specification provides conditions of the target system that are to be verified by the abstract interpretation. The specification may be represented as error states. The refinement system checks whether an abstract interpretation satisfies its specification by performing the steps of the abstract interpretation. The refinement system performs the first step of the abstract interpretation on the initial state to generate resulting states. The refinement system then applies a widening operator to the resulting states to establish over-approximated states of the first step. The refinement system then
performs the second step of abstract interpretation and applies the widening operator to the resulting states of the second step to establish an over-approximation of the states of the second step. The refinement system continues this process for the third and subsequent steps. After the states of a step are established, the refinement system determines whether the states of the step include an error state and thus will not satisfy the specification. If the states include an error state, the refinement system determines whether the error is a false error or a true error. To make this determination, the refinement system identifies which of the previous steps established a state as a result of the over-approximation of the widening operator that led to the error state. If the refinement system does not identify such a step, then the error is a true error. If the refinement system does identify such a step, the refinement system then applies a more precise operator at that step to establish new states for that step that do not include the state that led to the error state. The refinement system then re-performs the next step using the new states to establish states and continues this process for subsequent steps using the widening operator until the states converge. If the refinement system subsequently detects that the states of a step include an error state and identifies a step whose widening operator led to the error state (e.g., the source of the imprecision leading to the error), the refinement system will establish the states of subsequent steps using the widening operator or using a more precise operator if the step had previously been identified as the source of an imprecision leading to the error. In this way, the refinement system identifies errors in an abstract interpretation, determines whether the error is a false error or a true error, and automatically refines the abstract interpretation to remove the false errors.
[0012] Figure 1 is a block diagram that illustrates components of the refinement system in one embodiment. The refinement system 100 includes a target program 101, a generate abstract interpretation component 102, and an abstract interpretation 103. The generate abstract interpretation component inputs the target program and its specification and generates the abstract interpretation for the target program. The refinement system also includes a check abstract interpretation component 104, an identify source component 105, a widening operator component 106, and a more precise operator component 107. The check abstract interpretation component inputs the abstract interpretation and indicates whether the abstract
interpretation of the target program satisfies the specification. The check abstract interpretation component performs the steps of the abstract interpretation using the widening operator component. When a step establishes a state that is an error state, the check abstract interpretation component invokes the identify source component to identify the step that is the source of the imprecision leading to the error. The check abstract interpretation component invokes the more precise operator component to more precisely calculate the states of that step. If the check abstract interpretation component does not identify the source of the imprecision leading to the error, then the component indicates a true error. If the check abstract interpretation component determines that the states of the steps of the abstract interpretation have converged, then the component reports that the abstract interpretation has been verified as satisfying the specification.
[0013] The computing devices on which the refinement system may be implemented may include a central processing unit, memory, input devices (e.g., keyboard and pointing devices), output devices (e.g., display devices), and storage devices (e.g., disk drives). The memory and storage devices are computer-readable media that may contain instructions that implement the refinement system. In addition, the data structures may be stored or transmitted via a data transmission medium, such as a signal on a communications link. Various communications links may be used to connect components of the system, such as the Internet, a local area network, a wide area network, a point-to-point dial-up connection, a eel! phone network, and so on.
[0014] Embodiments of the refinement system may be implemented in various operating environments that include personal computers, server computers, multiprocessor systems, microprocessor-based systems, network PCs, minicomputers, mainframe computers, distributed computing environments that include any of the above systems or devices, and so on. The user computing devices may include cell phones, personal digital assistants, smart phones, personal computers, programmable consumer electronics, digital cameras, and so on.
[0015] The refinement system may be described in the general context of computer-executable instructions, such as program modules, executed by one or more computers or other devices. Generally, program modules include routines, programs, objects, components, data structures, and so on that perform particular
tasks or implement particular abstract data types. Typically, the functionality of the program modules may be combined or distributed as desired in various embodiments.
[0016] Figure 2 is a flow diagram that illustrates the processing of the check abstract interpretation component of the refinement system in one embodiment. The component is passed an abstract interpretation and returns an indication of whether the specification can be verified. In blocks 201-208, the component loops performing each step of the abstract interpretation until the states converge or an error is detected. In block 201, the component selects the next step for the abstract interpretation. In decision block 202, if the previous steps have already converged on a solution (e.g., current states are the same as the previous states), then the component completes with an indication that the specification has been verified, else the component continues at block 203. In block 203, the component establishes the states for the selected step using a widening operator (or more precise operator as appropriate). In decision block 204, if the established states include an error state, then the component continues at block 205, else the component loops to block 202 to select the next step. In block 205, the component invokes the identify source component, which performs precise backward reachability starling from the discovered error states to identify the source of the imprecision leading to the error. In decision block 206, if the error is a false error, then the component continues at block 207, else the component completes indicating a true error. In block 207, the component indicates to use a more precise operator for the step that is the source. In block 208, the component selects the source step as the next step and then loops to block 203 to start the re-establishing of the states of the steps.
[0017] Figure 3 is a flow diagram that illustrates the processing of the identify source component of the refinement system in one embodiment. The component is passed a step that generated an error state and identifies the source of the imprecision that leads to the error or indicates that the error is a true error. In block 301, the component initializes the suspect states to the detected error state. The suspect states are those states that might have led to the detected error state. In blocks 302-306, the component loops selecting previous steps to determine which step was the source of the error. In block 302, the component selects the previous step starting with the passed step. In decision block 303, if the states of the
selected step equal the initial state, then the error is a true error because no widening operator led to the error state and the component returns an indication of a true error, else the component continues at block 304. In block 304, the component identifies states of the selected step that led to the suspect states. In decision block 305, if some states are identified, then the selected step is not the source of the imprecision leading to the error and the component continues at block 306, else the component returns an indication that the error is a false error along with an indication of the step that is the source of the false error. In block 306, the component sets the suspect states to those that led to the current suspect states. The component then loops to block 302 to determine whether the previous step is a source of the error state.
[0018] The refinement system operates on a possibly infinite domain called "States." The domain States has a precise LUB operator u and a widening operator V. A transition system 0, representing a computer program, is a pair (1,6} where Ic States and θ: 2states →2states. The operator e is referred to as the "image" operator, which takes a set of states as input and gives the set of possible next states as output. The operator d~l is referred to as the "pre-image" operator, which takes the set of states as input and gives a set of possible previous states as output.
[0019] The refinement system represents programs as transition systems. A program P is a triple (V,I,T) where
V is a finite set of variables, each of which takes a value from a potentially infinite domain. For example, a program may have the variable count that has the domain of all integers and the variable index that has the domain of 1 to 100. A state represents a set of values for all the variables in V . For example, (count- 201, index = 50) and (count - 300, index = 1) are both states. The set of all possible values of V is the domain States, For example, the set of every possible combination of count and index represents the domain of States.
-I is a set of initial values of variables in v .
- Tc States x States is a binary relation such that T(s,s') holds
whenever it is possible for the program to transition from state s to state s in one step. For example, if index is only incremented by one in the program, then it is not possible to transition from a state with index = 5 to a state with index = 7 in one step.
A program P =(V,l,T ] thus provides a transition system 0 = {/,#}, where 0(S) = {*'|3j6 ST(s,./)}, and £T' (S)*[s 3/6 S.r(s,s'}}.
[0020] The refinement system represents a set of error states of a program as a specification PC States. To check if a transition system Q = {/,#) satisfies a
specification p (i.e., does not transition to an error state), the refinement system computes an over-approximation to the set of reachable states of the system, and checks if the over-approximation intersects {e.g., does not satisfy) specification p. If the over-approximation does satisfy the specification, then an exact or precise set of reachable states would also satisfy the specification. The refinement system represents that exact set {or least fixpoint) as PreciseReach(0) = //x./uXu#(X) (where µX.A represents the least fixpoint computation of expression A), though the fixpoint computation may not terminate. The transition system 0 satisfies specification v if and only if PreciseReach(0)np = 0, that is, the precise set includes no error states.
[00211 The refinement system uses a widening operator V to help ensure termination of a fixpoint computation at the cost of losing precision. If St and S2 are each a set of states such that s, cS2, then S3 =^V52 is a set of states such that 5, c 53 and S2 c .S3. Further, there is some metric (such as the number of conjunctions in the formula representing the set) that decreases from S, to S3. When the refinement system uses an over-approximation or a least fixpoint idenReach(e) = ^X.(/uX)V(/uXu<9{X)), (1) the computation of
WidenReach() will terminate, and (2) WidenReach(0)a PreciseReach(e). Thus, the transition system 0 will satisfy the specification » when
WidenReach(0) = 0, that is, when it does not include any error states, If, however, WidenReach(0)n^#0, the refinement system cannot distinguish
between the possibilities that either the transition system 9 does not satisfy specification p or the computation of WidenReach lost too much precision.
[0022] The refinement system keeps track of the intermediate states in the fixpoint computation, so that it can generate an abstract counterexample that can be automatically analyzed to classify if the error found is a false error or a true error. If it is a false error, the refinement system can also identify the precise point or step at which the abstract counterexample needs to be refined to avoid the recurrence of the specific false error.
[0023] The refinement system thus represents the fixpoint computation by the following equation:
The fixpoint computation can be further represented by the following equation:
where Rf represents the state of the transition system 0 at step / and K0 = / . If n is the smallest index of a step in which an error is detected (i.e., /?nnj!>*0), the refinement system looks from step n backwards to find the step whose widening is the source of the imprecision leading to the false error. If (/?„_, u0(fln_,))np = 0,
then step n is the source of the false error. To prevent precision loss at this step, the refinement system replaces the widening operator with the IUB operator in that particular step of the fixpoint computation. Otherwise, if (/?,,_, u #(#„_, ))np*0, then
the refinement system needs to backtrack to the source of the imprecision leading to the error. The refinement system identifies suspect states as #, ~ 9'1 (Rn CMP) , that
is, error states that are in Kn . Any step that has one of the suspect states may be
the source of the imprecision leading to the error. The refinement system then checks whether the states of the previous step satisfy the specification. If so, the widening system refines the widening operator of the next step n-1 to make it more precise. If not, then the widening of the previous step did not result in step n not
satisfying the specification. The refinement system continues this process until it finds either an index where the widening operator needs to be refined into a LUB operator (u) to avoid the false error, or the repeated backward propagation of the error state intersects with the initial state Ro = I , in which case the error is a true error.
[0024] The refinement system represents an abstract domain D = (D, -,0,®} as a join-semilattice where D is a concrete domain, |- is a partial order, o is the bottom element of the lattice, and the LUB dl @d, exists for all pairs of elements in D, that is, dltd2eD. For example, convex hull is such an operator for convex polyhedra. For all t/,,J2 e D , dt -d, means that dt is ordered strictly before dt = d, . The set of finite non-redundant subsets of D is ^(D,j-). A reduction function
QJ, ':p, . (D) -»#),-„ (D) maps each finite set into its non-redundant counterpart as represented by the following equation:
{d e S \d = 0 v 3d'e 5.J J-S, if and only if W, e 5, / 3d2 e S2.J, \-d2; that is, S1 is followed by S2 if and only if each state of S1 leads to a state in S, .
[0027] The refinement system represents the relation S1 p. S2 to be true if and only if either 5, = 0,, or S, -,, S2 and W2 e 5, / 3dl e Srd\ |-d, . The refinement system defines a connector operator □ such that for all SvSle[pfn(D,\-}, if 5, -fS2 , then S1p(S1□S2). Intuitively, the refinement system obtains S1□S2, by minimally combining the elements of S2 so as to obtain an S2, such that S1pS2. More precisely, if S2 is a maximal subset of S2 such that Vde5, :3d,e £,.OAm = 0)v(,x;0). The new symbolic state that is
generated after one more execution of the loop body is Sluw=(x 0 A m := 0) v (* £ A7 A x = w +1 A m > 0) v (* < AT A * = m + 2 A m > 0)
[0029] The refinement system computes S3 @S'2▼ps'2 Since S2pS'2 does not
hold, the refinement system computes S1□S2 by merging some of the elements of
S2. In particular, the refinement system merges
the refinement system merges ^f with the first element of s', then the result is S.; S3 5' @(x > 0 A m > 0 A A- ^ m). Otherwise, if the refinement system merges S2 with the second element of S2, then the result is 52' S3 s" @(x > 0 A m = 0) v (x < N A m > 0 A x > m +1 A x < m + 2). The result of widening in the first case is A->OAW>O, which is less precise than the widening result of the second case (,V>OAW = O)V(Am + lAm>0) when using the widening,
operator defined in Cousot and Haibwachs for the base domain of convex polyhedra. Thus, choosing the minimal result (second case) obtains a fixpoint that is strong enough to prove the assertion.
[0030] The widening operator ▼p satisfies the convergence properties of a widening operator, Without transforming S2 to S'2 using the S3 connector, such
convergence guarantees cannot be given. The operator u described above corresponds to the ®P operator.
[0031] For powerset abstract domains like the sets of convex polyhedra, the least upper bound operator ® is the non-redundant union. Thus, the refinement system will add more and more disjuncts to the reachable set of states. It is possible that this increase in the number of disjuncts will continue infinitely even though the assertion can be satisfied by an abstract fixpoint computed by merging some intermediate disjuncts and doing widening later on. Thus, intermediate merging may provide convergence. In one embodiment, the refinement system uses the connector EB operator as an operator to merge some disjuncts into one convex polyhedra. The refinement system checks whether using the merging operation instead of widening avoids error. If it does, then the refinement system replaces the widening operator by the merge operation. Thus, the refinement system has three upper bound operators u, □, and ▼ of decreasing precision. The refinement system can refine a widening operator to either the □ operator or the u operator. It can also refine the EB operator to the u operator. The hints that are generated by the refinement system as described below are of the form (i,op), where i is the step
number and ope {u, □ ] is the operator to be applied after that step.
[0032] If the refinement system returns (i,u), then it is clear that refining the widening operator in step i to LUB will remove the abstract counterexample. However, if the refinement system returns (i, □ }, then the widening operator in step i could be replaced either by a u or by □, it is not cfear whether it is desirable to convert the widening operation to u or □ in this case. It is possible that exactly one choice results in computation of the abstract fixpoint necessary to prove the property, whereas the other choice leads to non-termination of the abstraction refinement cycle. Thus, it may be more advantageous to try both possibilities.
[0033] The refinement system also addresses the non-monotonicity of widening. Widening is non-monotonic; if 5,cS', and S1£S'2, then it is not
necessarily the case that (S1▼S2)<(S1▼S2). Thus, refining the widening operator to a least upper bound operation in step i of the abstract fixpoint computation could result in a larger set of states in a later iteration. The refinement system avoids this problem of a larger set of states by, at every iteration of the abstract fixpoint computation, intersecting the states reached at step ; with the set of states reached at step i during the previous iteration of the abstract fixpoint computation. If the step count i is greater than the number of steps of the previous iteration, then the refinement system intersects the states with the states the last fixpoint computed in the previous iteration. The handling of non-monotonicity is described in Gulavani, B., and Rajamani, S., "Counterexample Driven Refinement for Abstract Interpretation," Microsoft Corp., MSR-TR-2006-02, January 2006, which is hereby incorporated by reference.
[0034] Figure 4 is a flow diagram that illustrates the processing of an abstract refine component of the refinement system in one embodiment. The abstract refine component is an implementation of the check abstract interpretation component as described above. The component is passed a transition system and a specification. The component uses variable hints to track steps that have been previously identified as the source of an imprecision leading to an error. In block 401, the component initializes the variable hints to be empty. In blocks 402-405, the component loops generating fixpoints until an error is detected. If the error is true, then the component completes, else the component identifies the step that is a source of the imprecision leading to the error and uses a more precise operator. In block 402, the component invokes an abstract fixpoint component to calculate the fixpoints for the transition system. In decision block 403, if the invoked component determines that the result of the fixpoint calculation indicates an error, then the component continues at block 404, else the component returns an indication of no error. In block 404, the component invokes a refine component to refine the widening of the source of the error. In block 405, the component adds the source of the imprecision leading to the error to the variable hints and then loops to block 402 to continue performing the steps of the abstract interpretation.
[0035] Figure 5 is a flow diagram that illustrates the processing of the abstract fixpoint component of the refinement system in one embodiment. The component is passed the transition system, the specification, and the hints. The component performs the steps of the abstract interpretation until an error is detected or it is determined that the transition system satisfies the specification. When the
component is first invoked, it initializes the state to the initial state and to the first type. In decision block 501, if the initial state is an error state, then the component returns, else the component continues at block 502. In decision block 502, if the current step was the source of an imprecision leading to the error state as indicated by the variable hints, then the component continues at block 503, else the component continues at block 504. In block 503, the component uses a more precise operator for calculating the states of the current step. In block 504, the component uses the widening operator to calculate the states of the current step. In decision block 505, if a state of the current step is an error state, then the component returns an indication of an error, else the component continues at block
506. In decision block 506, if the states of the current step has the same states as a
last step, then the abstract interpretation has been verified as satisfying the
specification and the component returns, else the component continues at block
507. In block 507, the component selects the next step and then loops to block 501
to process that step.
[0036] Figure 6 is a flow diagram that illustrates the processing of a refine component of the refinement system in one embodiment. The component is passed the transition system and the specification. The component locates the step that is the source of the imprecision leading to the error and indicates whether it is a false error or a true error. In block 601, the component sets the current step to the step that generated the error state. In decision block 602, if not all the previous steps have already been selected, then the component continues at block 603, else the component reports a true error in block 607 and completes. In block 603, the component sets the suspect states to the previous suspect states that are in the states of the current step. In decision block 604, if the states of the previous step and the states that those states led to do not include a suspect state, then the component returns an indication that the current step is a source of the error, else the component continues at block 605. In block 605, the component selects the previous step. In block 606, the component sets the suspect states to the states that led to the current suspect states and then loops to block 602.
[0037] From the foregoing, it will be appreciated that specific embodiments of the invention have been described herein for purposes of illustration, but that various modifications may be made without deviating from the spirit and scope of the
invention. Accordingly, the invention is not limited except as by the appended claims.

CLAIMS
I/We claim:
[c1] 1. A method in a computer system for determining whether an error
that has occurred during abstract interpretation of a transition system is a true error or a false error, the transition system specifying an initial state, transitions between states, and an error state, the method comprising:
detecting that a current step leads to the error state; and upon detecting that a current step leads to the error state, repeating the following until a false error or a true error is detected, when the union of states of the previous step of the current step and states that are directly led to by the states of the previous step includes no current suspect state that may lead to the error state, indicating a false error; setting new current suspect states to the states that directly lead
to the current suspect states; setting the current step to the previous step; and when the states of the current step include only the initial state, indicating a true error,
[c2 ]2. The method of claim 1 including upon detecting that a current step leads to the error state, setting the initial suspect states to the error state,
[c3] 3. The method of claim 1 wherein the transition system specifies
transitions from a current state to a next state for each state.
[c4] 4, The method of claim 1 wherein the transition system is derived
from a computer program.
[c5] 5. The method of claim 1 including establishing states for a current
step based on states of a previous step starting from the initial state in accordance with the transition system.
[c6] 6. The method of claim 5 wherein the establishing of states includes
using a widening operator to over-approximate the states.
[c7] 7. The method of claim 6 including when a false error is indicated,
establishing the states of the current step more precisely than when previously established using the widening operator.
[c8] 8. The method of claim 7 including re-establishing the states for
steps after the current step at which the false error was indicated.
[c9] 9. The method of claim 7 wherein the re-establishing of states takes,
for each step of re-establishing, the intersection of new states for the step with the states of a previous iteration for that step.
[c10] 10. A method in a computer system for performing abstract
interpretation of a program represented by a transition system, the transition system specifying an error state, the method comprising:
establishing states for steps of the transition system using a widening
operator that over-approximates the states; and upon detecting that the established states for a step includes an error state, identifying one of the previous steps that included a state that led
to the error state as a result of the over-approximation; establishing states for the identified step using an operator that more precisely specifies the states than the widening operator; and
establishing states for steps after the identified step based on the re-established states of the identified step,
[c11] 11. The method of claim 10 including when the over-approximation of
the widening operator did not result in the error state, indicating that the program includes an error.
[c12] 12. The method of claim 10 wherein the transition system specifies
transitions from one state to another state for each state,
[ciaj 13. The method of claim 10 wherein an initial state is specified by the
transition system.
[c14] 14. The method of claim 10 wherein the more precise specification of
the states is based on a least upper bound.
[c15] 15. The method of claim 10 including recording the identified step so
that subsequent re-establishing of the states of the identified step can also be more precise,
[c16] 16. The method of claim 10 wherein the identifying includes
backtracking through the steps using suspect states starting with the error state until a step whose states do not include a suspect state is found.
[ci7] 17. The method of claim 16 wherein the identified step is the step
after the step whose states do not include a suspect state.
[c18] 18. The method of claim 16 wherein during backtracking the suspect
states are set to the states of the previous step that led to the current suspect states.
[c19] 19. A computer system for performing abstract interpretation based
on a transition system, the transition system specifying an initial state, transitions between states, and an error state, comprising:
a component that, for each step specified by the transition system, establishes states of the steps based on the transition system using a widening operator that over-approximates states and using a more precise operator when the widening operator previously led to a step with one of its states including the error state; and
a component that detects whether states for a step include an error state, identifies one of the steps with a state that led to the error state because of the widening operator, indicates that the states for the identified step are to be established using the more precise operator, and directs the component that establishes states to establish states starting at the identified step.
[c20] 20. The computer system of claim 19 wherein when no step is
identified, indicating that the transition system includes a true error.

T