Efficient Program Instrumentation

Abstract

A method for determining the number and location of instrumentation probes to be inserted into a program is disclosed. The method advantageously inserts the minimum number of probes that are required to obtain execution coverage for every node in the program's control-flow graph. In addition, the method requires only type of node marking and one bit to store each probe, and does not require the assignment of weights to arcs or nodes of the control-flow graph. In the illustrative embodiment, the nodes of a control-flow graph are partitioned into non-empty sets, where each non-empty set corresponds to a super nested block of the program.

Description

FIELD OF THE INVENTION

The present invention relates to run-time analysis of software in general, and, more particularly, to a method of program instrumentation.

BACKGROUND OF THE INVENTION

Instrumentation is a technique that can enable engineers to comprehend, monitor, and assess the operation of software. Typically, a program is instrumented by inserting probes at various points in the program, where the probes report a variety of information such as whether certain portions of a program have been reached (referred to as coverage), the number of times that various portions of the program have been executed (referred to as execution counts), how much time is spent in various portions of the program, and so forth. Instrumentation thus facilitates the identification of coverage efficiency, bottlenecks, bugs, and other deficiencies in a program and, consequently, can aid in the process of improving the quality, security, efficiency, and performance of programs.

The introduction of probes into a program, however, adds overhead that can slow down the execution of the program, and thus there is a tradeoff when inserting probes into a program. Ideally, the probes should cover all of the various execution paths of the program, and should be sufficient in number so that the reported information is fine-grained enough to be useful. However, if there are too many probes, then program runtime performance might suffer appreciably, which is unacceptable in applications such as real-time embedded systems and Voice over Internet Protocol (VoIP).

Typically methods for determining probe insertion points in a program are based on a control-flow graph that is derived from the program. FIG. 1 depicts illustrative program 100, and FIG. 2 depicts control-flow graph 200 corresponding to program 100, both in accordance with the prior art. As shown in FIG. 2, control-flow graph 200 comprises nodes 201-1 through node 201-13, connected by arcs as shown. For convenience, each node of control-flow graph 200 has been assigned a label that indicates the portion of program 100 (known as a basic block) to which it corresponds.

In one method of the prior art, known as a maximum spanning tree method, arcs are first added to the control-flow graph, as necessary, so that at each node, the incoming execution count equals the outgoing execution count. Weights are then assigned to the arcs of the (possibly-augmented) control-flow graph, and a maximum spanning tree is generated (i.e., a spanning tree such that the sum of its arc weights is maximum.) Finally, a probe is inserted at every node in the control-flow graph that leads to an arc not in the spanning tree.

FIG. 3 depicts an illustrative maximum spanning tree for control-flow graph 200, indicated by boldface arcs, in accordance with the prior art. (For simplicity, weights are not depicted in the figure.) As shown in FIG. 3, an arc from node 201-13 to node 201-1 has been added to ensure that the incoming and outgoing execution counts are equal at each node.

It is readily apparent from FIG. 3 that the following arcs are not part of the spanning tree:

• • (201-6, 201-8) [B-E6],
  • (201-9, 201-11) [D-E6E]
  • (201-4, 201-12) [F-E1E]
  • (201-12, 201-2) [E1E-E1]
  • (201-2, 201-13) [E1-G]Consequently, probes are inserted in nodes B, D, F, E1E, and E1.

A key disadvantage of the maximum spanning tree method is that it requires execution counts on each probe, which can consume a great deal of memory. Moreover, the counter values can grow so large that they impact the original application, and there is no way to reset the counters. Consequently, the maximum spanning tree method is typically not practical for program monitoring during field operation.

In another method of the prior art, known as a super block dominator method, a pre-dominator tree of the control-flow graph is first generated—i.e., a tree in which a first node is an ancestor of a second node if and only if the first node is guaranteed to execute before the second node. FIG. 4 depicts pre-dominator tree 400 for control-flow graph 200, in accordance with the prior art.

Next, a post-dominator tree of the control-flow graph is generated—i.e., a tree in which a first node is a descendent of a second node if and only if the first node is guaranteed to execute before the second node. FIG. 5 depicts post-dominator tree 500 for control-flow graph 200, in accordance with the prior art.

The pre-dominator and post-dominator trees are then combined into a single dominator graph. FIG. 6 depicts dominator graph 600 for control-flow graph 200, in accordance with the prior art. Dominator graph 600 is simply the union of pre-dominator tree 400 and post-dominator tree 500, and can be obtained by adding the arcs of post-dominator tree 500 to pre-dominator tree 400.

Next, the strongly-connected components of the dominator graph are determined. A strongly-connected component is a maximal set of nodes in a directed graph such that every node in the set is reachable from every other node in the set. FIG. 7 depicts the strongly-connected components of dominator graph 600, in accordance with the prior art.

Finally, each strongly-connected component is defined as a respective super block, and a probe is inserted in each of the super blocks. In this example, a probe is inserted into each of the following super blocks of program 100: {A, E1, G}, {E2, E1E}, {F}, {E3, E6, E6E}, {B}, {C}, {D}, and {E}.

SUMMARY OF THE INVENTION

The present invention provides a novel method for determining the number and location of instrumentation probes to be inserted into a program. In particular, the illustrative embodiment advantageously inserts the minimum number of probes that are required to obtain execution coverage for every node in the program's control-flow graph. Moreover, the method requires only one type of node marking and one byte to store information for each probe. The illustrative embodiment thus minimizes instrumentation overhead that can slow down program execution, and is therefore especially well-suited to monitoring Voice over Internet Protocol (VoIP) software.

In accordance with the illustrative embodiment, the nodes of a control-flow graph are partitioned into non-empty sets, where each non-empty set corresponds to a super nested block of the program. A super nested block is a block of code such that for any two consecutive lines of code X and Y, if X executes, then Y also executes at some point after the execution of X, albeit possibly with one or more other lines of code executed in between X and Y. Thus, a super nested block might have a branching statement and correspond to a plurality of nodes of the control-flow graph, or instead might be a basic block and consist of a single node of the control-flow graph.

The illustrative embodiment employs a recursive method that identifies the set of super nested blocks while traversing a control-flow graph. The method can be performed on a control-flow graph that has already been derived from a program, or it can advantageously be performed while the control-flow graph is itself being constructed during parsing of the program. Once the super nested blocks of a program have been determined, a probe is inserted into each innermost layer of basic blocks. The outer-layer blocks' coverage information can be inferred from those probes. The resulting instrumentation enables execution coverage information to be obtained for every node and arc in the control-flow graph, with a minimum number of probes.

The illustrative embodiment comprises: partitioning a program into one or more blocks of code, wherein any two consecutive lines of code X and Y of said program are placed in the same block if and only if the execution of X implies the execution of both X and Y, albeit not necessarily consecutively; and inserting a probe into each of said blocks of code; wherein the partitioning of said program is based on a partition of the nodes of a control-flow graph G for said program; and wherein the nodes of said control-flow graph G are partitioned by an algorithm that employs a single type of mark to mark nodes.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 depicts an outline of illustrative program 100, in accordance with the prior art.

FIG. 2 depicts a control-flow graph corresponding to illustrative program 100, in accordance with the prior art.

FIG. 3 depicts an illustrative maximum spanning tree of control-flow graph 200, as shown in FIG. 2, for a first instrumentation method of the prior art.

FIG. 4 depicts a pre-dominator tree for control-flow graph 200 for a second instrumentation method of the prior art.

FIG. 5 depicts a post-dominator tree for control-flow graph 200 for a second instrumentation method of the prior art.

FIG. 6 depicts a dominator graph for control-flow graph 200 for a second instrumentation method of the prior art.

FIG. 7 depicts the strongly-connected components of dominator graph 600, as shown in FIG. 6, for a second instrumentation method of the prior art.

FIG. 8 depicts the high-level architecture of a first illustrative embodiment of the present invention.

FIG. 9 depicts a flowchart of the Super Nested Block Method, in accordance with the illustrative embodiments of the present invention.

FIG. 10 depicts a data-processing system for instrumenting programs in accordance with the first illustrative embodiment of the present invention.

FIG. 11 depicts the salient contents of memory 1020, as shown in FIG. 10, in accordance with the first illustrative embodiment of the present invention.

FIG. 12 depicts the high-level architecture of a second illustrative embodiment of the present invention.

FIG. 13 depicts a flowchart of the salient tasks performed by off-line analyzer 810, testing tool 830, auto generator 1201, compiler 1202, and run-time instrumenter 1220, as shown in FIGS. 8 and 12, in accordance with the second illustrative embodiment of the present invention.

FIG. 14 depicts a data-processing system for instrumenting programs in accordance with the second illustrative embodiment of the present invention.

FIG. 15 depicts the salient contents of memory 1420, as shown in FIG. 14, in accordance with the first illustrative embodiment of the present invention.

DETAILED DESCRIPTION

FIG. 8 depicts the high-level architecture of a first illustrative embodiment of the present invention. As shown in FIG. 8, the first illustrative embodiment comprises off-line analyzer 810, run-time instrumenter 820, and testing/monitoring tool 830, interconnected as shown.

Off-line analyzer 810 comprises software, or hardware, or a combination of software and hardware capable of determining one or more locations in a program at which an instrumentation probe is to be inserted. The determination of instrumentation locations by off-line analyzer 810—referred to as the Simplified Super Nested Block Method—is described in detail below. The method is first described informally as applied to illustrative control-flow graph 200, and subsequently a formal specification of the method is provided.

Informal Description of the Simplified Super Nested Block Method

An informal description of the Simplified Super Nested Block Method as applied to illustrative control-flow graph 200 is first provided in this section. A formal specification of the method is subsequently provided in the following section.

In the first task of the method, the first layer of super nested block starting with the root node (i.e., node 201-1 or “A” for control-flow graph 200) is identified. In the case of control-flow graph 200, the first layer of super nested block consists of node 201-1 (A), node 201-2 (E1), node 201-12 (E1E), and node 201-13 (G), which corresponds to the following lines of code:

Non-branching statementA1;
Non-branching statementA2;
......
Non-branching statementAn;
While Expression1 {
}
Non-branching statementG1;
Non-branching statementG2;
......
Non-branching statementGs;
}

As will be appreciated by those skilled in the art, after reading this specification, the lines of code above constitute a super nested block, because for any two consecutive lines of code X and Y in the block, if X is executed, then Y is also executed at some point after the execution of X, albeit possibly with one or more other lines of code executed in between X and Y. This first super nested block is subsequently referred to as SNB1.

The second task of the method checks whether the current super nested block (at this point, SNB1) has any branching statements. If not (i.e., the super nested block comprises a single node of the control-flow graph, and is thus simply a basic block), the single node is marked “probe-needed”. Otherwise, expand the branching points into the roots of a second-layer (or “child”) super nested block.

In the case of control-flow graph 200, the branching point E1 is expanded into the node 201-3 (E2) and the node 201-13 (G). Node 201-13 is a basic block itself and does not need further expansion and thus is marked as “probe-needed”. Node 201-3 is a branching point and needs further expansion.

In the case of control-flow graph 200, the second layer of super nested block is the first-layer statements inside the while loop, starting from node 201-3. This second super nested block, SNB2, consists of a single node, node 201-3 (E2), which corresponds to the following lines of code:

If Expression2 {
......
}
else {
......
}

The second task is then repeated for super nested block SNB2. Because SNB2 includes a branching statement, it is further expanded into a third layer with two super nested blocks. The first third-layer super nested block, SNB31, consists of node 201-5 (E3), node 201-8 (E6), and node 201-11 (E6E) and corresponds to the following lines of code:

Switch Expression3 {
...
}
If Expression6 {
...
} else {
...
}

Because super nested block SNB31 has more than one branching statement, all of them must be further expanded. The node 201-8 (E6) is expanded into node 201-9 (D) and node 201-10 (E) and the node 201-5 (E3) is expanded into nodes 201-6 (B) and 201-7 (C).

All four nodes, 201-9 (D), 201-10 (E), 201-6 (B), and 201-7 (C) must be marked with “probe-needed” because they cannot be further expanded. Note that even though the coverage of nodes 201-8 (E6) and 201-11 (E6E) can be determined by nodes 201-6 (B) and 201-7 (C), the coverage of nodes 201-9 (D) and 201-10 (E) cannot be determined by nodes 201-6 (B) and 201-7 (C), and thus both nodes 201-9 (D) and 201-10 (E) need to be selected in addition to nodes 201-6 (B) and 201-7 (C). This is a point where coverage is different from execution counters, where node 201-9 (D) can be determined by B+C−E or E can be determined by B+C−D.

The second third-layer super nested block, SNB32, consists of node 201-4 (F), and corresponds to the following lines of code:

Non-branching statementF1;
Non-branching statementF2;
......
Non-branching statementFr;

Because super nested block SNB32 does not have any branching statements (i.e., it is a basic block), node 201-4 (F) is marked “probe-needed”. The second task is now completed.

In the third and final task, a probe is inserted into the portions of source program 100 that correspond to the nodes marked “probe-needed”: node 201-6 (B), node C, node 201-9 (D), node 201-10 (E), and node 201-4 (F).

As will be appreciated by those skilled in the art, the expansion of subsequent layers of the control-flow graph lends itself very well to a recursive implementation, and this is in fact how the method is formally specified in the subsequent section. As will further be appreciated by those skilled in the art, in some other embodiments of the present invention the expansion might be performed in an alternative, non-recursive, fashion (e.g., iteratively via a breadth-first search traversal of the control-flow graph, etc.), and it will be clear to those skilled in the art, after reading this disclosure, how to make and use such alternative embodiments.

Formal Specification of the Simplified Super Nested Block Method

FIG. 9 depicts a flowchart of the salient tasks of the Simplified Super Nested Block Method, in accordance with the illustrative embodiments of the present invention. In the illustrative embodiments of the present invention, the method of FIG. 9 is performed by off-line analyzer 810.

At task 910, variable S is initialized to a singleton set containing the root node.

Task 920 checks whether there is a node V in S and a node W in G-S such that execution of the last line of code of V implies execution of the first line of code of W. If so, execution proceeds to task 930, otherwise execution continues at task 940.

At task 930, node W is added to set S.

Task 940 checks whether at least one node of set S has a branch statement. If so, execution proceeds to task 950, otherwise execution continues at task 960.

At task 950, the method is performed recursively for every child node of set S. After task 950, execution continues at task 970.

At task 960, one node of set S is marked as “probe-needed”.

At task 970, the program corresponding to control-flow graph G is modified to count the number of times that each node marked “probe-needed” is executed, or simply check if the node is executed at least once. After task 970, the method of FIG. 9 terminates.

Run-time instrumenter 820 comprises software, or hardware, or a combination of software and hardware capable of instrumenting program P during program P's execution, in well-known fashion. In accordance with the first illustrative embodiment of the present invention, run-time instrumenter 820 replaces each instrumentation location in the executing program (e.g., executable program P, etc.) with code patches for saving registers, running probes, restoring registers, removing probes after the first time that they are executed, and putting back the original code at the instrumentation location. In addition, run-time instrumenter 820 uses the probes to generate and report code coverage information concerning the execution of program P. In some embodiments of the present invention, run-time instrumenter 820 might report code coverage information after the execution of program P has completed, while in some other embodiments, run-time instrumenter 820 might report code coverage information during the execution of program P, while in still some other embodiments run-time instrumenter 820 might report code coverage information both during and after the execution of program P.

The automatic generated run-time instrumenter includes two parts: the static creation and dynamic parameter setting. The static part includes the following program portions: 1) create a patch object as the one single instance of the top-level class; 2) create a new process of the program under testing or have a running one attached; 3) create a probe; 4) set instrumentation points; 5) insert the probe to the instrumentation points; 6) repeat parts 3), 4), and 5) until all instrumentation points are properly handled. The dynamic part is the memory size and value of instrumentation location.

Testing/monitoring tool 830 comprises software, or hardware, or a combination of software and hardware capable of running a run-time instrumenter (e.g., run-time instrumenter 820, etc.) and an executable under test (e.g., executable program P, etc.) in parallel, in well-known fashion. During the testing of the target program, whenever the instrumentation point is reached, the execution is redirected to the code patches of saving registers, running probes, restoring registers, and restoring the original code back to the instrumentation point.

FIG. 10 depicts data-processing system 1000 for instrumenting programs in accordance with the first illustrative embodiment of the present invention. As shown in FIG. 10, data-processing system 1000 comprises processor 1010 and memory 1020, interconnected as shown.

Processor 1010 is a general-purpose processor that is capable of executing instructions stored in memory 1020, of reading data from and writing data into memory 1020, and of executing the tasks associated with off-line analyzer 810 and run-time instrumenter 820, as described above. As will be appreciated by those skilled in the art, in some alternative embodiments of the present invention, processor 1010 might instead be a special-purpose processor; in any case, it will be clear to those skilled in the art, after reading this disclosure, how to make and use processor 1010.

Memory 1020 stores data, program source code, and executable instructions, as is well-known in the art, and might be any combination of random-access memory (RAM), flash memory, disk drive, etc. In accordance with the first illustrative embodiment of the present invention, memory 1020 stores the source code for a particular program P to be instrumented, the executable instructions (i.e., object code) for program P, an executable program for performing the tasks of off-line analyzer 810, and an executable program for performing the tasks of run-time instrumenter 820, as shown in FIG. 11.

FIG. 12 depicts the high-level architecture of a second illustrative embodiment of the present invention. As shown in FIG. 12, the second illustrative embodiment comprises off-line analyzer 810 and testing tool 830 of the first illustrative embodiment, as well as auto generator 1201, compiler 1202, and run-time instrumenter 1220, interconnected as shown.

Auto generator 1201 comprises software, or hardware, or a combination of software and hardware that is capable of generating source code for a run-time instrumenter based on (i) the source code for program P, and (ii) the instrumentation locations determined by off-line analyzer 810. In accordance with the second illustrative embodiment of the present invention, auto generator 1201 generates source code for the run-time instrumenter that is in the same programming language as program P. The auto generator first generates a template of the code, and then replaces the dynamic portion, memory size and probe locations with actual value calculated from the analysis step. The code is output in the same programming language as the original program under testing/monitoring.

Compiler 1202 comprises software, or hardware, or a combination of software and hardware that is capable of generating an executable program from source code, in well-known fashion.

Run-time instrumenter 1220 is an executable software program capable of instrumenting program P during program P's execution, in well-known fashion. In accordance with the second illustrative embodiment of the present invention, run-time instrumenter 1220 replaces each instrumentation location in program P with code patches for saving registers, running probes, restoring registers, removing probes after the first time that they are executed, and putting back the original code at the instrumentation location. In addition, run-time instrumenter 1220 reports code coverage information concerning the execution of program P. In some embodiments of the present invention, run-time instrumenter 1220 might report code coverage information after execution of program P has completed, while in some other embodiments, run-time instrumenter 1220 might report code coverage information during the execution of program P, while in still some other embodiments run-time instrumenter 1220 might report code coverage information both during and after the execution of program P.

FIG. 13 depicts a flowchart of the salient tasks performed by off-line analyzer 810, testing tool 830, auto generator 1201, compiler 1202, and run-time instrumenter 1220, in accordance with the second illustrative embodiment of the present invention.

At task 1310, off-line analyzer 810 determines instrumentation locations for program P in accordance with the method of FIG. 9, as described above.

At task 1320, auto generator 1201 generates source code for run-time instrumenter that is in the same programming language as program P, based on the program P source code and the instrumentation locations determined at task 1310, as described above.

At task 1330, compiler 1202 compiles the program P source code and run-time instrumenter source code, generating a program P executable and run-time instrumenter 1220, in well-known fashion.

At task 1340, testing tool 830 executes program P and run-time instrumenter 1220 in parallel, in well-known fashion.

After task 1340 is completed, the method of FIG. 13 terminates.

FIG. 14 depicts data-processing system 1400 for instrumenting programs in accordance with the second illustrative embodiment of the present invention. As shown in FIG. 14, data-processing system 1400 comprises processor 1410 and memory 1420, interconnected as shown.

Processor 1410 is a general-purpose processor that is capable of executing instructions stored in memory 1420, of reading data from and writing data into memory 1420, and of executing the tasks associated with off-line analyzer 810, auto-generator 1201, compiler 1202, and run-time instrumenter 1220, as described above. As will be appreciated by those skilled in the art, in some alternative embodiments of the present invention, processor 1410 might instead be a special-purpose processor; in any case, it will be clear to those skilled in the art, after reading this disclosure, how to make and use processor 1410.

Memory 1420 stores data, program source code, and executable instructions, as is well-known in the art, and might be any combination of random-access memory (RAM), flash memory, disk drive, etc. In accordance with the second illustrative embodiment, memory 1420 stores the source code for a particular program P to be instrumented, the executable instructions (i.e., object code) for program P, an executable program for performing the tasks of off-line analyzer 810, the auto-generated run-time instrumenter source code, and executable run-time instrumenter 1220, as shown in FIG. 15.

It is to be understood that the disclosure teaches just one example of the illustrative embodiment and that many variations of the invention can easily be devised by those skilled in the art after reading this disclosure and that the scope of the present invention is to be determined by the following claims.

Claims

1. A method comprising: partitioning a program into one or more blocks of code, wherein any two consecutive lines of code X and Y of said program are placed in the same block if and only if the execution of X implies the execution of both X and Y, albeit not necessarily consecutively; andinserting a probe into each of said blocks of code;wherein the partitioning of said program is based on a partition of the nodes of a control-flow graph G for said program; andwherein the nodes of said control-flow graph G are partitioned by an algorithm that employs a single type of mark to mark nodes.

2. The method of claim 1 wherein said partition of the nodes of said control-flow graph G is determined during the generation of said control-flow graph G.

3. The method of claim 1 wherein the probe insertion is performed during the generation of said control-flow graph G.

4. The method of claim 1 further comprising: determining a non-empty subset S of nodes of said control-flow graph G;partitioning G-S into one or more subgraphs that correspond to child nodes of said subset S; anddetermining a partition of each of said subgraphs.

5. The method of claim 4 wherein the determination of said non-empty subset S comprises solving a minimum-cut maximum-flow problem.

6. The method of claim 4 wherein the determination of subgraph partitions is performed recursively.

7. The method of claim 1 wherein said control-flow graph G is unweighted.

8. A method comprising: marking one or more nodes of a control-flow graph G of a program with a single type of mark during the generation of said control-flow graph G; andinserting one or more probes into said program during the generation of said control-flow graph G based on the node markings.

9. The method of claim 8 further comprising partitioning the nodes of said control-flow graph G during the generation of said control-flow graph G such that any two consecutive lines of code X and Y of said program are in the same partition if and only if the execution of X implies the execution of both X and Y, albeit not necessarily consecutively.

10. The method of claim 9 wherein the partitioning comprises: determining a non-empty subset S of nodes of said control-flow graph G;partitioning G-S into one or more subgraphs that correspond to child nodes of said subset S; anddetermining a partition of each of said subgraphs.

11. The method of claim 10 wherein the determination of said non-empty subset S comprises solving a minimum-cut maximum-flow problem.

12. The method of claim 10 wherein the determination of subgraph partitions is performed recursively.

13. The method of claim 8 wherein said control-flow graph G is unweighted.

14. A method comprising: (1) identifying a maximal set S of nodes of a control-flow graph G such that: (a) said set S comprises a unique root node of said control-flow graph G; and(b) if (i) a node V belongs to said set S, and(ii) said control-flow graph G has an arc from said node V to a node W, and(iii) execution of the last line of code of said node V implies that the first line of code of said node W is also subsequently executed, albeit not necessarily consecutively,thensaid node W also belongs to said set S;(2) when at least one node of said set S comprises a branching statement, performing said method recursively for every child node of said set S, wherein in each recursive invocation of said method a child node functions as a new root node for a new control-flow graph, and wherein said new control-flow graph is a proper subset of said control-flow graph G; otherwise, marking a node of said set S; and(3) inserting a probe into a program that corresponds to said control-flow graph G at points in said program corresponding to each marked node.

15. The method of claim 14 wherein said method is performed during the generation of said control-flow graph G.

16. The method of claim 14 wherein the identification of said set S comprises solving a minimum-cut maximum-flow problem for said control-flow graph G.

17. The method of claim 14 wherein when the root node of said control-flow graph G is unmarked, said method further comprises: estimating a potential execution count for one or more child nodes of said set S, and determining said proper subset of child nodes of said set S based on the estimation.

18. The method of claim 14 wherein said control-flow graph G is unweighted.