JP2682206B2 - Optimization method of target program - Google Patents

Description

The present invention relates to an optimization method of a target program, and particularly to a target program for inputting an atomic program described in a high-level programming language to a multiprocessor system and performing parallel processing. The optimization method of the object program in the compiler that generates

[Conventional technology]

Conventionally, in this type of compiler, in the process of inputting a source program to a multiprocessor system and generating a target program for parallel processing, parallelization is performed before and after cache clear processing required for synchronous or exclusive control. When data that has become non-shared data is defined or quoted by, the instructions can be rearranged even if the instructions can be rearranged so that cache clear processing is not performed in the middle of a series of definitions or quotation instructions for the data. Didn't.

FIG. 2 is a diagram showing an example of the source program 100 written in FORTRAN. This source program 100 includes an external subroutine (subroutine name: PARALLEL) 110 that is executed in parallel processing.
In addition, an assignment statement 120 that defines a local variable A that becomes non-shared data by parallelization, an external subroutine (subroutine name: BARRIER) call statement 130 for synchronization, and a local variable An assignment statement 140 quoting A is included in this order.

In addition, FIG. 3 shows the source program 10 illustrated in FIG.
FIG. 8 is a diagram showing intermediate text 200 generated by a conventional compiler for 0. This intermediate text 200 includes an assignment statement 200 that defines a local variable A, an external subroutine call statement 230 for synchronization, and an assignment statement 240 that cites a local variable A in this order. ing. When the external subroutine call statement 230 for synchronization is executed by executing the assignment statement 220 that defines the local variable A, the local variable A exists in the cache, but by execution of the external subroutine call statement 230 Since the cache is cleared, a cache miss will occur in the assignment statement 240 that immediately cites the local variable A, which makes it impossible to execute the program at high speed.

[Problems to be solved by the invention]

In the conventional compiler described above, in the process of inputting the atomic program to the multiprocessor system and generating the target program for parallel processing, parallelization is performed before and after the cache clearing process required for synchronous or exclusive control. When data that has become non-shared data is defined or quoted, the instructions are rearranged even if they can be rearranged so that cache clear processing is not performed in the middle of a series of definitions or quotation instructions for the data. Since it was not, if the data that has become non-shared data is defined or quoted before the cache clear processing and if the data is defined or quoted after the cache clear processing, after the cache clear processing There is a cache miss in the definition or quotation of There is a drawback that it is not possible to run without the program at a high speed.

In view of the above points, an object of the present invention is to define or quote data that has become non-shared data due to parallelization before and after cache clear processing required when performing synchronization or exclusive control, and An object of the present invention is to provide an optimization method of a target program in which instructions are rearranged so that cache clear processing is not performed in the middle of a series of definitions or quoted instructions.

[Means for solving the problem]

The optimization method of the object program of the present invention is to input a source program written in a high-level programming language to a multiprocessor system and generate an object program for parallel processing. And a parallelization processing unit for generating a first intermediate text, a parallelization processing unit for inputting the first intermediate text, and generating a second intermediate text by parallelizing parallelizable portions in a program, and the second intermediate text. And a non-shared data extraction unit that outputs information about data that has become non-shared data by parallelization to a non-shared data table, and the second intermediate text and the non-shared data table are input to perform synchronization or exclusive control. Defines or pulls the data that became non-shared data before and after the cache clear process required when performing And an optimization processing unit that generates a third intermediate text in which the second intermediate text is rearranged so that cache clear processing is not performed in the middle of a series of definitions or citation commands for the data. A target program generation unit for inputting the intermediate text and generating a target program.

[Action]

In the optimization method of the object program of the present invention, the syntactic analysis unit inputs the source program and performs syntactic analysis to generate the first intermediate text, and the parallelization processing unit inputs the first intermediate text and parallelizes the program. Second parallelization of possible parts
The intermediate text is generated, the non-shared data extraction unit inputs the second intermediate text, outputs information about the data that has become non-shared data by parallelization to the non-shared data table, and the optimization processing unit outputs the second intermediate text. When defining or quoting data that has become non-shared data before and after the cache clearing process that is required when inputting the non-shared data table and performing synchronization or placement control, a series of definitions or quotation commands for the data are defined. The third intermediate text is generated by rearranging the second intermediate text so that the cache clear processing is not performed in the middle, and the target program generation unit generates the third intermediate text.
Input intermediate text and generate target program.

〔Example〕

Next, the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a compiler 2 to which an object program optimization method according to an embodiment of the present invention is applied. This compiler 2 is the source program 1
Is input to perform parsing to generate the first intermediate text 3, and the parallel to input the first intermediate text 3 and generate the second intermediate text 4 in which the parallelizable parts in the program are parallelized. The non-shared data extraction unit 23 that inputs the second intermediate text 4 and outputs information about the data that has become non-shared data by parallelization to the non-shared data table 5, and the second intermediate text 4 When defining or quoting data that has become non-shared data before and after the cache clearing process necessary for inputting the shared data table 5 and performing synchronization or exclusive control, a series of definition or citation commands for the data An optimization processing unit 24 that generates a third intermediate text 6 in which the second intermediate text 4 is rearranged so that the cache clearing process is not performed on the way.
And a target program generation unit 25 for inputting the third intermediate text 6 and generating the target program 7.

FIG. 4 is a diagram showing an example of the non-shared data table 5 generated by the shared data extraction unit 23 for the source program 100 illustrated in FIG. In the non-shared data table 5, a non-shared data name is stored as information about the data that has become non-shared data.

Next, the operation of the object program optimizing method of the present embodiment configured as described above will be described.

The syntactic analysis unit 21 inputs the source program 1 and performs syntactic analysis to generate the first intermediate text 3.

Next, the parallelization processing unit 22 inputs the first intermediate text 3 and generates the second intermediate text 4 in which the parallelizable parts in the program are parallelized.

Then, the non-shared data extraction unit 23 inputs the second intermediate text 4 and outputs information about the data that has become non-shared data by parallelization, for example, a non-shared data name to the non-shared data table 5.

Next, the optimization processing unit 24 inputs the second intermediate text 4 and the non-shared data table 5 and collects the data that has become non-shared data before and after the cache clear processing required when performing synchronization or exclusive control. When the definition or reference is made, the third intermediate text 6 in which the second intermediate text 4 is rearranged is generated so that the cache clear processing is not performed in the middle of a series of definition or reference commands for the data.

FIG. 5 is a diagram showing the third intermediate text 6 generated by the optimization processing unit 24 for the source program 100 illustrated in FIG. Here, the assignment statement 340 that cites the local variable A includes only the data (A and C) whose assignment statement 340 has become non-shared data based on the non-shared data table 5 by the optimization processing unit 24. Since it is handled, it is determined that it is possible to move to the external subroutine call statement 330 for synchronization, and the local variable A is set in front of the external subroutine call statement 330 for synchronization. Since there is the assignment statement 320 to be defined, it is determined to move immediately before the external subroutine call statement 340 for synchronization in consideration of the cache miss at the time of data definition or reference, and the second intermediate text 4 is set to the third. It is rearranged like the intermediate text 6.

Finally, the target program generator 25 inputs the third intermediate text 6 and generates the target program 7.

Therefore, the target program 7 generated by the compiler 2 for the source program 100 illustrated in FIG.
Is executed, a cache miss does not occur in the assignment statement 340 that cites the local variable A, so that the program can be executed at high speed.

〔The invention's effect〕

As described above, according to the present invention, in a compiler that inputs a source program to a multiprocessor system and generates a target program that performs parallel processing, a non-process before and after cache clear processing required when performing synchronization or exclusive control is performed. When data that has become shared data is defined or quoted, by rearranging the instructions so that the cache clear processing is not performed in the middle of a series of definition or quote instructions for the data, the Even when the data that has become the shared data is defined or quoted and the data is defined or quoted after the cache clear processing, the cache clear processing is performed in the middle of a series of definition or quote commands for the data. The instructions are reordered so that they do not occur, so the cache clear Since the cache miss when the instruction which has been defined or cited the data after processing is executed does not occur, there is an effect that the program can be executed at high speed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a compiler to which an object program optimization system according to an embodiment of the present invention is applied, FIG. 2 is a diagram showing an example of a source program written in FORTRAN, and FIG. Is a diagram showing an intermediate text generated by a conventional compiler for the source program illustrated in FIG. 2, and FIG. 4 is a first diagram for the source program illustrated in FIG.
The figure which shows the non-shared data table produced | generated by the non-shared data extraction part in the figure, FIG. 5 is the 1st with respect to the source program illustrated in FIG.
It is a figure which shows the 3rd intermediate text produced | generated by the optimization process part in a figure. In the figure, 1 ... Source program, 2 ... Compiler, 3 ... First intermediate text, 4 ... Second intermediate text, 5 ... Non-shared data table, 6 ... Third intermediate text, 7 ... Purpose Program, 21 …… Syntax analysis unit, 22 …… Parallelization processing unit, 23 …… Unshared data extraction unit, 24 …… Optimization processing unit, 25 …… Object program generation unit, 100 …… Primitive program, 110… … External subroutine executed in parallel, 120,220,320 …… Assignment statement that defines local variable A, 130,230,330 …… External subroutine call statement for synchronization, 140,240,340 …… Assignment statement that cites local variable A Is.

Claims (1)

(57) [Claims] 1. A compiler for inputting an atomic program written in a high-level programming language to a multiprocessor system to generate an object program for parallel processing, inputting a source program and performing syntax analysis to output a first intermediate text. A parsing unit for generating, a parallelization processing unit for inputting the first intermediate text, and for generating a second intermediate text by parallelizing parallelizable portions in a program, and inputting the second intermediate text for parallelization. A non-shared data extraction unit that outputs information about the data that has become non-shared data to a non-shared data table, and a cache that is required when performing synchronization or exclusive control by inputting the second intermediate text and the non-shared data table When defining or quoting data that has become non-shared data before and after clear processing An optimization processing unit for generating a third intermediate text in which the second intermediate text is rearranged so that the cache clearing process is not performed in the middle of a series of definitions or quotation commands for the target program for inputting the third intermediate text. And a target program generation unit that generates a target program optimization method.