JPH04124737A - Data dependence analyzing system - Google Patents

Abstract

PURPOSE:To easily analyze the data dependence by making a data dependence analysis interrupting means collect the trace information at execution of a multi-task program after conversion and compare this trace information with the serial trace information and then interrupting the analysis of the data dependence when the data dependence is detected. CONSTITUTION:A trace information collecting means 7 carries out a serial program 5 after its conversion and obtains the serial trace information 8 based on an inserted trace row. Thus the information 8 is obtained with execution of the program 5. Meanwhile a data dependence analysis interrupting means 9 collects the trace information on a multi-task program 6 based on the inserted trace row when the program 6 is carried out after its conversion. Then the trace information on the program 6 is compared with the collected information 8 for analysis of the presence or absence of the data dependence. That is, it is analyzed whether the emerging sequence of definition/quotation of variables (array elements with array) is equal to the value of the emerging sequence or not. If not, the presence of dependence is confirmed and the execution of analysis of the data dependence is interrupted with addition of a prescribed message.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a method for analyzing data dependence of a multitask program in an electronic computer system.

[Conventional technology]

Multiprocessor systems are often aimed at improving throughput and turnaround time in multi-programming environments, but they can also be used to decompose a program into multiple tasks and process them in parallel. It is also an effective system for improving processing speed. In particular, when processing speed is important, such as in supercomputers, performance is often improved through multitasking.

For this reason, if you want to run a serial program that used to run on a single processor system on a multiprocessor system, you need to rewrite it to a multitasking program with a structure that makes effective use of the multiprocessor and can be executed at high speed. will be held.

By the way, when rewriting a serial program to a multitask program, you should be careful that if there is a data dependency between tasks, the order of data definitions and citations may be reversed compared to the serial program. There are cases where malfunction occurs and correct results are not obtained.

Therefore, in order to prevent malfunctions in advance, it is common to analyze the dependencies of multitask programs as part of the debugging process. Conventionally, the following two methods have been mainly adopted.

■Method of statically analyzing source programs ■Method of taking and examining time-series execution traces for each task [Problem to be solved by the invention] As mentioned above, two methods have been mainly adopted in dependency analysis in the past. However, each had the following drawbacks.

First, with the method of statically analyzing a source program in (■), it is often difficult to thoroughly examine all definitions and citation relationships for each element of an array, especially when the control flow including procedure calls is complex. It becomes very difficult to analyze dependencies in these cases.

On the other hand, with the method (2) of examining time-series execution traces for each task, it is difficult to narrow down the events to be traced, and tracing many events results in a huge amount of trace data, making it difficult to analyze dependencies. It becomes difficult to extract the essential trace data from there.

As described above, it is not easy to perform dependency analysis using either method.

The present invention has been proposed in view of the above points, and its purpose is to provide a data dependency analysis method that allows easy dependency analysis.

[Means for Solving the Problems] In order to achieve the above object, the present invention inputs a serial program and a multitask program equivalent thereto, and collects trace information about variables specified by inter-task shared variable specification information. a program conversion means for inserting a trace line to be executed; a trace information collection means for collecting serial trace information by executing a converted serial program; A data dependency analysis interrupting means for interrupting execution when data dependency is detected by comparison is provided.

[Effect]

In the data dependency analysis method of the present invention, the program conversion means inputs a serial program and an equivalent multitask program, and performs a trace line for collecting trace information about variables specified by inter-task coexistence variable specification information. is inserted, the trace information collection means executes the converted serial program and collects serial trace information, and the data dependency analysis interrupting means collects trace information when executing the converted multitask program, and collects the serial trace information and the serial trace information. Interrupts execution when data dependencies are detected during comparison.

[Example] Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the data dependency analysis method of the present invention.

In FIG. 1, this embodiment includes a program converting means 4, a trace information collecting means 7, and a data dependency analysis interrupting means 9 as functional parts. The functions and other parts of each means will be explained through the following operations.

An operator prepares a serial program 1 and a multitask program 2 and supplies them to the program conversion means 4. Here, serial program 1 is a program that was executed in a single processor system, and multitask program 2 is an equivalent program for a multiprocessor system that has been rewritten from serial program 1 by a programmer using an appropriate tool.

On the other hand, the operator provides the inter-task shared variable designation information 3 to the program conversion means 4 from a predetermined input device. FIG. 2 shows an example of inter-task shared variable specification information 3, which specifies that dependency analysis is to be performed on variable (array) A in common block C0M.

The program conversion means 4 inputs both the serial program 1 and the multitask program 2, recognizes parallel execution sections, etc., and then converts the variables to be subjected to dependency analysis specified by the inter-task shared variable specification information 3. Insert a trace line (instruction to collect trace information) in the reference line of
A converted serial program 5 and a converted multitask program 6 are output. Note that the following trace information is appropriate to collect.

・If the variable is an array, the array element number ・Reference location (program unit name 1 statement number) ・Reference format (defined II + EF)
, quotation "REF")・Variable value However, if there is no statement number in the original program, the program converting means 4 assigns an appropriate value to the statement number of the reference location.

The above program conversion will be explained below using a specific example.

FIG. 3 is an example of a serial program 1 written in the FORTRAN language, and FIG. 4 is an example of an equivalent multitask program 2. Note that the multitask program 2 in this example is created on the assumption that parallel processing will be performed by two processors. That is, the change range of the Do variable at the outermost circumference of each Do small loop in the serial program 1 in FIG. 3 is divided into two, and two tasks are configured so that each is processed in parallel.

Now, assuming that the inter-task shared variable specification information 3 specifies the variable A as already shown in FIG. Enter task program 2,
A converted serial program 5 as shown in FIG. 5 and a converted multitask program 6 as shown in FIG. 6, in which a trace line is inserted next to the reference line of variable A, are output.

That is, when a description of variable A appears, the program conversion means 4 determines whether its reference format is a definition or a quotation, and if there is no statement number, it assigns an appropriate statement number, and converts the Do variable. value.

Subroutine TRA for collecting trace information using reference format, program unit name 9 statement number, and value of variable A as arguments
Insert a CALL statement that calls CE and DEPEND.

Note that for the multitask program 2, parallel execution sections are recognized in advance in order to determine the program unit name. In FIG. 4, this recognition can be determined from CALL FORK indicating the start part of parallel processing, CALL JOIN indicating the end part of parallel processing, and the task ID specified by those arguments. It is recognized that and section P2 are processed in parallel.

Next, in FIG. 1, the trace information collecting means 7 executes the converted serial program 5 and obtains serial trace information 8 based on the inserted trace line. Fifth
By executing the serial program 5 after conversion shown in the figure, the seventh Tj! Serial trace information 8 having the content shown in J is obtained. In addition, in FIG. 7, "displacement" is an array element number.

On the other hand, in FIG. 1, the data dependency analysis interrupting means 9 collects trace information of the converted multi-task program 6 based on the inserted trace line when executing the converted multi-task program 6, and collects trace information of the converted multi-task program 6 based on the inserted trace line. The presence or absence of data dependence is analyzed by comparing it with the serial trace information 8 that is included.

In other words, it analyzes whether the order of appearance and values of definitions and citations of variables (in the case of arrays, each array element) are the same, and if a difference occurs, it is determined that there is a dependency and a predetermined message is sent. to interrupt execution.

FIG. 8 shows an example of trace information obtained when executing the converted multitask program 6 of FIG. 6. According to the serial trace information 8 of FIG. 7, array element A (30) is quoted with statement number "20" after it is defined with statement number "90001J" and the value is "50", but in the trace information in Figure 8, statement number "20" is defined before array element A (30) is defined. 20", and its value is also "0", which is different from the correct one. Therefore, in this case, it is determined that there is data dependency, and execution is interrupted at this point.

That is, FIG. 9 shows the multitasking program of FIG. 4 again, but with 5LIBROLITINE
Quotation of variable A in LINE8 of SUB and PRO
This means that there is a dependency with the definition of variable A in LINE 33 of GRAM MAIN.

Further, FIG. 10 shows an example of a message output when dependency is detected.

Although the debugger was not mentioned in the above example, the multitasking program 6 after conversion is run under the debugger.
It is also possible to perform the debugging process efficiently by allowing the debugging process to be started when a data dependency is detected and the execution is interrupted.

〔Effect of the invention〕

As explained above, the data dependency analysis method of the present invention has the following features: (1) Since it is based on dynamic analysis, it is possible to search for dependencies without omission even when the control flow is complex. I can do it.

(2) Just by specifying inter-task shared variables, if a dependency is detected from the trace results, it will be known, so you can narrow down the events to be traced based on the huge amount of trace information! : There is no need to consider it, and the number of debugging steps can be significantly reduced.

There are other effects.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an example of the data dependency analysis method of the present invention, FIG. 2 is a diagram showing an example of inter-task shared variable specification information, FIG. 3 is a diagram showing an example of a serial program, Figure 4 shows an example of a multitasking program, Figure 5 shows an example of a serial program after conversion, Figure 6 shows an example of a multitasking program after conversion, and Figure 7 shows an example of serial trace information. Figure 8 is a diagram showing an example of trace information of a multitask program after conversion, Figure 9 is a diagram showing dependencies between tasks, and Figure 1O is a message output when dependency is detected. FIG. In the diagram, l... Serial program 2... Multitask program 3...:...
Inter-task shared variable specification information 4...Program conversion means 5...Conversion serial program 6...
・Multitask program after conversion 7... Trace information collection means 8... Serial trace information 9... Block diagram of an embodiment of data dependency analysis interrupting means Diagram: Sharing between tasks FIG. 1 is a diagram showing an example of variable specification information; FIG. 2 is a diagram showing an example of a PROGRAM MAIN serial program. PROGRAM MAIN COMMON /COM/ N INITA, INITB UNRomatic diagram showing an example of a multitasking program PROGRAM MAIN Figure showing an example of a serial program after conversion NITBND

Claims (2)

[Claims] (1) A program conversion means that inputs a serial program and an equivalent multitask program and inserts a trace line for collecting trace information about variables specified by inter-task shared variable specification information, and executes the serial program after conversion. trace information collection means for collecting serial trace information when the multitasking program is executed after conversion; 1. A data dependency analysis method characterized by comprising a means for interrupting gender analysis. (2) The data dependency analysis method according to claim 1, wherein the trace information to be collected is an array element number, a reference location, a reference format, and a value of a variable for an array.