JPH02163833A - Block dividing method and synchronizing process position deciding method for program repeating structure and program converting device - Google Patents

Abstract

PURPOSE:To obtain a synchronizing process position in a simple procedure by dividing a program repeating structure into blocks where the only input/ output flow exists respectively in the case the repeating structure includes a complicated control structure and this structure has the data dependeint relation. CONSTITUTION:A program analyzing part 11 analyzes the branching destination of each sentence, the data using way, the additional characters, etc., for a sequential program to be converted. Based on this analyzed data, a synchronizing process position deciding part 12 decides a synchronizing process position between loops. Then a code generating part 13 sets a synchronizing instruction into the synchronizing process position decided at the part 12 to produce a parallel processing code. Thus the synchronizing process position is obtained in a simple procedure.

Description

[Detailed Description of the Invention] Industrial Application Field The present invention relates to a block division method for dividing a repetitive structure of a program into blocks having only one input/output flow, and a method for parallel processing of loops having data dependencies. The present invention relates to a synchronous processing position determination method for determining the execution position of a synchronous instruction, and a program conversion device for converting a sequential program into a program that is executed while maintaining synchronization at locations having data dependencies.

BACKGROUND OF THE INVENTION Conventional program conversion devices have a method of inhibiting parallelization when there is a data dependency between processes executed in parallel, or determining the execution position of a synchronized instruction between processes.
Select all branch positions that lead to or can be reached from the synchronization processing position.

Problems to be Solved by the Invention In conventional program conversion devices, even if the execution position of a synchronous instruction can be determined, the number of synchronous instructions is large because the synchronous instruction is executed every time there is a branch. Furthermore, there is no way to make this decision in simple steps. Another problem is that the procedure for optimizing multiple synchronization processes is complicated.

In view of the above problems, the present invention aims to obtain a synchronous processing position by a simple procedure when converting a sequential program into a parallel processing program, and furthermore,
It is an object of the present invention to provide a synchronous processing position determination method used in a device and a block division method therefor.

Means for Solving the Problems In order to solve the above problems, the block division method of the present invention includes the steps of generating a flow graph, depth-first search, and division into blocks with a unique flow of people. The synchronous processing position determining method includes the steps of dividing the program, initializing the synchronous processing position, and removing redundant synchronous processing.

The program conversion device is comprised of the steps of program analysis, determining the synchronous processing position, and generating code for parallel processing.

Effect When converting a sequential program into a parallel program using the program conversion device configured as described above, if synchronization processing is required due to data dependencies between loops, the synchronization instruction is used to ensure that there is only one input/output flow. Since it is executed after mI of the divided block, the number of instructions is small and easy to understand.

Furthermore, since the block division method is used to determine the execution position of the synchronization command, the execution can be performed in a simple procedure.

Furthermore, block division can be performed in a simple procedure, making it suitable for depth-first search.

EXAMPLES The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the overall configuration of an embodiment of the program converting device of the present invention, which is composed of a program analysis section 11, a synchronous processing position determination section 12, and a parallel code generation section 13.

In FIG. 1, a program to be converted of a sequential type is first analyzed by a program analysis section 11 to determine the branch destination of each statement, how data is used, and its subscripts. Based on the data analyzed here, the synchronization processing position determination unit 12 determines the synchronization processing position between loops. The code generation unit 13 generates a code for parallel processing that incorporates a synchronization command at the synchronization processing position determined by the synchronization processing position determination unit 12.

FIG. 2 is a flowchart showing details of an embodiment of the synchronization processing position determination method of the present invention, which is comprised of a process 21, a synchronization processing position initial determination 22, redundant synchronization removal 23, and the like. In FIG. 2, the program analysis section 1 in FIG.
If there is a data dependency relationship between the loops, the loop analyzed in step 1 is divided into blocks having only one input/output flow in a block division step 21. Next, in a step 22 for initial determination of synchronous processing position, the synchronous processing position is initially determined. In this method, a block is used as a unit of synchronization processing, and the synchronization processing positions are set before and after a block that includes a statement that requires synchronization processing. Finally, in the step 23 of removing redundant synchronization, if there is unnecessary synchronization due to multiple synchronizations, this is removed.
For example, this can be realized by the procedure shown in FIG.

FIG. 4 is a flowchart showing details of an embodiment of the block division method of the present invention, which includes flow graph creation 41, depth-first search 42, block division 43, and the like. In FIG. 4, when the loop shown in FIG. 5 is divided into blocks, a flow graph as shown in FIG. 6 is created in a flow graph creation step 41. Next, in step depth-first search 42, a depth-first search is performed on the flow graph of FIG. 6 to find the Lot of each sentence.

DFNUMBKR and the backedge of the search tree are determined as shown in FIG. Here, DFNUMBlle
R(:n) indicates the visiting order of sentence n, and Lot(n) is
This is the smallest DFNUMBER of sentences that can be returned by tracing backward from sentence n (branches that were not passed when reaching the final sentence in depth-first search). moreover,
In a step called block division 43, the search tree of FIG. 7 is divided into blocks as shown in FIG.

This can be achieved by regarding the sentence n, DFNUMBKR(n)=LOWCn), as the beginning of each block.

Effects of the Invention Since the present invention is configured as described above, it produces the effects described below.

By using depth-first search to divide a program into blocks, it becomes possible to easily divide the program into blocks that have only one input/output flow.

By using block division to determine the position of synchronization processing between loops having data dependence, it becomes possible to easily determine the position, including the removal of redundant synchronization.

Furthermore, by using the above-mentioned synchronous processing position determining method for converting a sequential program into a parallel program, a program conversion device that can relatively easily obtain a parallel program with fewer synchronous processing positions can be realized.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the overall configuration of an embodiment of the program conversion device of the present invention, FIG. 2 is a flowchart showing details of an embodiment of the synchronous processing position determination method of the present invention, and FIG. FIG. 4 is a flowchart showing details of an embodiment of the block division method of the present invention; FIG.
The figure shows an example of the target program, Figure 6 is a flow graph created for the loop in Figure 6, and Figure 7 is when depth-first search is performed on the flow graph in Figure 6. FIG. 8 is a block division diagram when the search tree of FIG. 7 is divided into blocks. 11...Program analysis department, 12...
Synchronous processing position determination unit, 13... Parallel code generation unit, 21... Block division, 22...
Initial determination of synchronization processing position, 23... Removal of redundant synchronization, 41... Creation of flow graph, 42.
...Depth-first search, 43...Flock division. Name of agent: Patent attorney Shigetaka Awano and 1 other person 1st
Figure 3 Figure beI+n Chi c1. - jam, m-1,..., 1f□r D
I, 2. ..., m 1+@1lln rf 5vII (+, ko) a+ ℃Pler+d. +or M, -, J Child v town..., m 5YN(1,k)-0 1st wi2 Figure 42 ■ Do 101~1. N ■ K婁2+L ■ Do 20 J-1, N C(f)-J εLSE C(1◆2)#TSUKASA10 cONTINuE LOW[+]”l■0FNL114BεI?[+], 1
10■8--1 13■f1: bac god dge Fig. 6 jar ■-■ + ◆ ■ ■ Fig. 8

Claims (3)

[Claims] (1) A block division method that divides the repetitive structure of a program into blocks having only one input/output flow when the repetitive structure of the program includes a complicated control structure and has data dependencies therein. (2) When dividing a repeating structure and processing it in parallel,
2. A synchronous processing position determining method for determining the execution position of a synchronous instruction between parallel processing processes based on data dependence using the block division method according to claim 1. (3) A program conversion device that uses the synchronous instruction position determination method according to claim 2 when converting a program written in a sequential language to one for parallel processing.